SenTrojan: Revealing Analog Hardware Trojan Attacks on Sensors

SenTrojan: Revealing Analog Hardware Trojan Attacks on Sensors Kun Wang1 , Ruochen Zhou2,* , Wenjun Zhu1 , Shilin Xiao1 , Kai Wang1 , Chen Yan1 , Xiaoyu Ji1 , and Wenyuan Xu1

arXiv:2611.06103v1 [cs.CR] 4 Nov 2026

1

Zhejiang University, China, {eewk, zwj , xshilin, eekaiwang, yanchen, xji, wyxu}@zju.edu.cn 2 The Hong Kong University of Science and Technology, China, zrccc@ust.hk

Abstract—Sensors bridge the physical and digital worlds, making their integrity critical to cyber-physical system security. PDoS Existing sensor attacks largely exploit unintended and sensorspecific physical vulnerabilities, limiting their generalizability and Trigger Signal Temperature motivating a different threat model, i.e., whether an adversary can Sound Pressure Static Vcc implant hidden attack surfaces into sensors before deployment. Vcc Spoofing Inserted In this paper, we present SenTrojan, an analog hardware Trojan Trojan framework for sensors. SenTrojan embeds malicious functionality Vin Dynamic Attacker into analog signal circuitry commonly shared across diverse sensor Spoofing architectures, and activates the hidden behavior through crafted Trojan-inserted Sensors Sensor Failure external physical signals. Once activated, SenTrojan can induce sensor malfunctions ranging from permanent denial-of-service to Fig. 1: An illustration of SenTrojan. The attacker can use a static or dynamic spoofing. To achieve these effects under different pre-designed trigger signal to activate the analog sensor Trojan, architectural constraints, we introduce two complementary designs. The Logic-Rule-Driven (LRD) Trojan adds lightweight analog leading to PDoS, static spoofing, or dynamic spoofing.Operating Trigger Signal monitoring logic to detect trigger features such as frequency and Normal Attacker amplitude, enabling reliable activation across various sensor types. The Physical-Effect-Driven (PED) Trojan targets sensors with vulnerabilities arise unintentionally Benign Sensors exploitable native analog effects, e.g., nonlinearity and rectification, However, because these and realizes attacks with minimal component insertion. We during design or implementation, they are often inconsistent evaluate SenTrojan on 14 sensors spanning 6 transduction across sensor models, hardware versions, and deployment principles and demonstrate its effectiveness across diverse real- environments. As a result, existing attacks typically require world scenarios. Overall, SenTrojan achieves an average successful pre-existing exploitable weaknesses, and substantial effort has trigger rate above 95%, maintains over 97% similarity to benign behavior, incurs zero false-trigger rate for LRD designs, and been devoted to characterizing and detecting them [6]. Rather than passively searching for existing sensor vulnerintroduces less than 5% average deviation in drift, noise floor, and response time. To our knowledge, this work is the first abilities, we investigate an alternative threat model, i.e., the systematic study of analog sensor Trojans, broadening sensor deliberate insertion of exploitable vulnerabilities into sensors security from unintended physical vulnerabilities to intentional before deployment. This threat is supported by real-world supply-chain compromise.

…

precedents and established digital paradigms, e.g., hardware Trojans [7]–[10] and neural network backdoors [11]. However, Sensors serve as the critical interface between physical this paradigm remains largely unexplored in analog sensor environments and digital systems, underpinning a wide range circuitry. Although prior work has explored Trojan designs of applications, including autonomous vehicles, industrial using analog components [12], these designs primarily replace automation, and healthcare monitoring. The reliability and digital counters with analog capacitors, while still monitoring safety of these systems depend fundamentally on correct sensor digital states and relying on specific data patterns or instructions operation. Catastrophic consequences resulting from sensor for activation. Their triggers are therefore ultimately softwarefailures, such as airplane crashes [1], automotive accidents [2], issued binary events rather than physical signals received and industrial disasters [3], underscore the importance of sensor through the sensing interface. This leaves open the central challenges of analog sensor Trojans, i.e., how to preserve the security. accuracy of the analog front end while achieving reliable Trojan Existing sensor security research has predominantly focused activation under noisy physical inputs. on exploiting inherent vulnerabilities through signal injection In this paper, we introduce SenTrojan, the first supplyattacks. These approaches leverage unintentional hardware imperfections, e.g., circuit nonlinearity [4] and electromag- chain analog hardware Trojan framework for sensors. Unlike netic (EM) susceptibility [5], to compromise sensor readings. prior sensor attacks that passively exploit pre-existing physical vulnerabilities, we examine an active compromise model in * Ruochen Zhou is the corresponding author. which an adversary deliberately inserts exploitable analog I. I NTRODUCTION

Sound Heat

Sensitive Element

Conversion Element

Signal Conditioning Circuits ADC

Amp.

UAV

CPS

Transducer

LPF

Smart Home

…

Physical Stimulus

Sensor Light

…

Fig. 2: The typical hardware structure of a sensor.

1 2

addresses this challenge by systematically analyzing ADC sensor modules and circuit characteristics to identify parameters that can be tuned without noticeably degrading normal operation. By tuning effects such as nonlinearity, rectification, and frequency-dependent coupling, PED Trojans convert crafted physical inputs into low-frequency output deviations. This allows SenTrojan to induce exploitable behavior in sensors resistant to known signal injection attacks, or amplify weak existing vulnerabilities with minimal component insertion. Our contributions are summarized as follows: • We present the first systematic study of analog sensor Trojans, where crafted physical signals activate malicious behavior without software privileges or digital access. • We design SenTrojan, a two-paradigm framework for analog sensor Trojans that combines broadly applicable logic-driven Trojans with minimally inserted effect-driven Trojans. • We implement and evaluate SenTrojan across 14 sensors spanning 6 transduction principles, demonstrating attack feasibility, stealthiness, and real-world impact.

…

behavior into the sensor circuitry before deployment. The attacker can activate the hidden Trojan using a crafted external physical signal, without requiring software privileges or digital access to the victim system, as shown in Fig. 1. SenTrojan expands the sensor attack surface in two ways. First, it enables attack effects that are difficult to achieve through direct signal injection alone, such as persistent output manipulation. Second, it can compromise sensors that are resistant to known signal injection attacks or amplify weak physical vulnerabilities that would otherwise have limited attack impact. To support these attack scenarios under different sensor architectures and insertion constraints, SenTrojan introduces two complementary designs. The Logic-Rule-Driven (LRD) Trojan adds lightweight analog circuitry to monitor crafted physical signals and control malicious payloads. By checking features such as frequency, amplitude, and duration against predefined rules, LRD Trojans provide reliable activation across diverse sensor types and support flexible payload behaviors. The Physical-Effect-Driven (PED) Trojan further targets sensors whose native analog effects can be induced or amplified by crafted inputs. PED Trojans avoid explicit monitoring logic and instead tune existing analog components so that the sensor’s own physical behavior serves as the trigger and payload mechanism. This design enables minimal hardware changes while inducing exploitable vulnerabilities in resistant sensors or strengthening weak existing vulnerabilities. However, realizing SenTrojan in the analog sensor domain requires addressing a common challenge of functionality and stealth, i.e., a Trojan must reliably respond to attackercontrolled physical triggers while preserving normal sensor behavior before activation. For LRD Trojans, the main challenge is reliable and selective triggering without perturbing the sensing path. LRD Trojans add explicit trigger and payload circuitry, but analog sensor signals are continuous, noisy, and highly sensitive to circuit loading. Additional circuitry may change path impedance, increase leakage current, or distort normal sensor outputs, while the trigger must still distinguish crafted physical signals from benign environmental variations. SenTrojan addresses this challenge by separating trigger monitoring from the native sensing path. It uses highimpedance parallel monitoring to observe signal features while minimizing loading effects during normal operation. To improve selectivity, LRD Trojans combine single-condition and multi-condition detection over features such as frequency and amplitude, together with temporal validation. Only when the crafted trigger satisfies the predefined rule for a sufficient duration does the Trojan activate the payload. For PED Trojans, the main challenge is inducing exploitable behavior without explicit trigger logic. Dedicated detection logic and payload modules increase hardware footprint [7], [13], especially in tightly constrained sensor circuits. PED Trojans therefore avoid explicit monitoring circuitry and instead rely on native analog effects in sensor components. However, without a separate trigger decoder, the Trojan must still transform crafted physical inputs into targeted sensor output corruption while preserving normal behavior under benign inputs. SenTrojan

II. BACKGROUND A. Sensor Structure A sensor is designed to generate electrical measurements in response to a physical stimulus [14], which is then transmitted to cyber-physical systems (CPS), such as unmanned aerial vehicles and smart homes, as illustrated in Fig. 2. A modern sensor comprises four main components: transducer, amplifier, filter, and analog-to-digital converter (ADC). The transducer is the core component of a sensor, consisting of a sensitive element and a conversion element. The sensitive element detects a specific physical stimulus, while the conversion element translates this stimulus into an analog electrical signal. Because the transducer’s output is generally weak and vulnerable to noise, the signal conditioning circuit incorporates amplifiers and filters to boost the signal and suppress noise. To make the signal suitable for digital processing, an ADC converts the analog output to digital form. B. Hardware Trojan Hardware Trojans are malicious circuits covertly embedded into hardware systems, such as additional or altered logic, that typically remain inactive until triggered by rare or specific conditions. The common methods to activate Trojans are specific digital inputs or software commands. Once activated, they can compromise security by leaking sensitive data or degrading system reliability. These Trojans pose a significant

2

threat, particularly in complex, multi-party semiconductor supply chains, because they can be inserted at various stages of the hardware lifecycle, including design, fabrication, and assembly [9], [15]. The design of a hardware Trojan can be decoupled into two parts: trigger and payload. The trigger determines under which conditions the malicious circuit becomes active, while the payload decides which malicious functionality the Trojan can achieve. Even though the trigger does not achieve any malicious functionality, it can largely improve the stealthiness of a hardware Trojan. An ideal trigger condition is nearly impossible to meet in the normal operation of the hardware, e.g., a random 1024-bit digital input.

after activation without continuous intervention and may require a system reset or sensor replacement for recovery. • Dynamic spoofing. The Trojan manipulates sensor readings into attacker-controlled time-varying signals. Unlike static spoofing, dynamic spoofing can produce measurements that may appear plausible to downstream systems. • Permanent denial-of-service (PDoS). The Trojan renders the sensor unable to provide valid measurements. This effect may result from permanent physical damage or from repeatedly forcing the sensor into a protection state. D. Design Requirement We consider that a successful SenTrojan attack should satisfy the following requirements: • Trojan functionality. The Trojan should be reliably activated by a predefined physical signal and induce the intended sensor malfunction after activation. • Trojan stealthiness. The Trojan should preserve normal sensor behavior before activation and avoid false activation under benign environmental stimuli, even under practical functional validation [12]. • Minimum components. The Trojan should use as few inserted or modified components as possible to reduce hardware overhead and signs of tampering.

III. T HREAT M ODEL A. Attacker Model

We consider an adversary who can compromise sensors before they are deployed or during later maintenance and assembly. This capability is consistent with modern sensor supply chains as shown in Appendix A, where design, fabrication, packaging, PCB assembly, and maintenance may be performed by different parties across organizational or geographic boundaries. Specifically, Trojans may be inserted at the design stage, where a malicious insider modifies circuit parameters or component choices [16]–[19]; the manufacturing stage, IV. D ESIGN where an untrusted foundry or fabrication process introduces This section details the SenTrojan design illustrated in Fig. 3. layout or process-level modifications [10], [12], [20]; or the Because attack execution requires the signal modality and trigpost-manufacturing stage, where a compromised assembler ger parameters to be tailored to each target sensor (Sec. IV-A), or maintenance adversary replaces components, adds small we develop a Trojan framework with two analog-hardware circuitry, or alters PCB traces [21]–[23]. The adversary may activation paths: an external logic-rule-driven (LRD) path be motivated by sabotage, espionage, or economic gain. Our and an internal physical-effect-driven (PED) path. LRD uses design does not depend on the insertion stage. It assumes only explicit high-impedance monitoring, feature validation, and a that the adversary can embed a Trojan in the sensor’s analog separate payload module, enabling flexible and controllable signal path without disrupting normal validation. Appendix B trigger-payload configurations across sensor types at the cost details the capabilities and limitations at each stage. of additional circuitry. In contrast, PED introduces or tunes analog physical effects, such as envelope detection, to minimize B. Attacker’s Capabilities component overhead and enable stealthy dynamic spoofing. We To avoid raising suspicion during inspection or functional detail the LRD and PED mechanisms and their circuit designs validation, the adversary is constrained in both the type in Secs. IV-B and IV-C, with emphasis on attack effectiveness and number of inserted components. We assume that the and stealth. adversary does not add digital modules, such as processors or microcontrollers, which are uncommon in sensor front ends A. Trigger Signal Design and require additional support circuitry. Instead, the adversary Trigger signal architecture necessitates precise modality may insert or modify only small-scale integrated (SSI) circuits, selection and parameter specification. e.g., logic gates and analog switches, and discrete components, Modality selection. Prioritizing practical long-range ine.g., resistors, capacitors, and diodes. These components are jection over close contact methods suggests four optimal common in sensor circuits and can be introduced with limited modalities: laser, sound, ultrasound, and EMI. For example, impact on the overall architecture. using ultrasound to trigger a Trojan is commonly considered more stealthy and practical than applying pressure to a pressure C. Attack Goal sensor. We exploit native transducer sensitivities to aforeThe adversary aims to compromise sensor measurements by mentioned signal modalities, such as accelerometer acoustic activating the implanted Trojan with a crafted physical signal. susceptibility [24] or microphone ultrasonic vulnerability [4]. We consider three representative attack goals. Alternatively, we integrate supplementary transducers, like • Static spoofing. The Trojan forces the sensor to output a photoresistors for optical responsiveness. A detailed comparison constant, predefined erroneous value. This effect can persist of these signal modalities is available in Appendix C.

3

Logic-Rule-Driven Trojan Design

Trigger Signal

Tested Signal (𝑉in )

Modality

Condition Detector

Frequency

Single-Condition Multi-Condition Frequency Combinational Amplitude Sequential

Amplitude

Physical-Effect-Driven Trojan Design Tested Signal (𝑉in )

𝑉in

t

𝑉𝑓𝑜𝑢𝑡 𝑇𝑑

Trojan Design

𝑉𝑓𝑜𝑢𝑡 𝑉 Activation 𝑑𝑜𝑢𝑡 Duration Detector Signal 𝑉𝑑𝑜𝑢𝑡 Charge Latch Payload Accumulation

LRD Trojan PED Trojan

t 𝑇𝑐

Amplifier Saturation Rectification Env-Det Over-Amp

Transducer Nonlinearity Coupling

𝑇𝑑 𝑇 𝑑

LPF+ ADC

Output Signal (𝑉𝑜𝑢𝑡 ) 𝑉𝑜𝑢𝑡

𝑉in

Normal

Filtered out 0 𝑉𝑜𝑢𝑡

𝑉𝑡h

t

PED Trojan

t

t

t

Fig. 3: An illustration of the SenTrojan design. Left: Trigger signal specification. Center: Logic-Rule-Driven (LRD) Trojans utilize discrete logic to validate signal features and duration (Td ) before latching the activation signal. Right: Physical-Effect-Driven (PED) Trojans leverage component physics like rectification to extract baseband payloads from high-frequency inputs, effectively bypassing protection filters to manipulate sensor measurements.

Peak Detector

HPF

or

Window comparator

Comparator Vcc

C1

4

Vfout

Vin

R1

R5

R1 C2

R2

R4

Vcc

Vfout

C1

R4

Vin

Vfout

R2

R3

Vp

Vf

R5

R1

Vin

C2

R3

(a) Frequency detector.

Comparator

Vcc

R3

Vp

Vf

Peak Detector

HPF

R2

R4

(a) Waveforms under 2 kHz input.

Vfout (b) Waveforms under 20 kHz input.

(b) Amplitude detector.

Fig. 5: Simulation waveforms of the frequency detector. Vin is the input, Vf is the filtered signal, and Vp is the peakdetector output. When Vp exceeds the threshold, the subsequent comparator is activated, triggering the payload.

Fig. 4: Schematics of a single condition detector.

Parameter design. Parameter optimization targets frequency coupling efficiency to extend attack range, requiring physical testing like sweeping. In terms of amplitude, higher signal Trojan triggering, as complex shape discrimination necessitates levels typically induce stronger interference. The design of prohibitive area and power overheads. As illustrated in Fig. 4, the Trojan activation mechanism necessitates a comparative the single-condition circuit translates true detection logic into a analysis of injected versus legitimate signal characteristics. high-level signal, denoted as Vf out , which activates the trigger Specifically, we characterize legitimate signals across their duration detector. datasheet-specified operational ranges to identify and exploit Frequency. To maintain high input resistance and preserve characteristic disparities as trigger thresholds. monitoring accuracy, our frequency triggering mechanism detects the preset frequency using a cascade of a frequency B. Logic-Rule-Driven Trojan Design detector, a peak detector, and a comparator, as depicted To prevent accuracy degradation, we employ a parallel in Fig. 4(a). Frequency detection can use LC resonance monitoring circuit that operates in a high-impedance state (frequency-selective behavior in an inductor-capacitor circuit) during quiescence, thereby preserving the native sensor preci- or high-pass filters (filters that pass frequencies above a cutoff sion with minimal signal perturbation. Furthermore, to evade frequency). Crucially, even if environmental noise coincides detection during brief factory functional tests [25], the Trojan with the trigger frequency, the subsequent comparator prevents incorporates a temporal activation threshold that exceeds false activation by enforcing a strict amplitude threshold. standard durations. LRD Trojan architecture integrates three Initially, the filter isolates the frequency range, setting cut-off core modules: the trigger condition detector, trigger duration frequencies, fp = 1/(2πR1 C1 ), above normal sensor operation detector, and payload. to mitigate non-ideal characteristics. An enhanced peak detector 1) Single-Condition Detector: Common characteristics of an subsequently minimizes adjacent frequency interference and analog signal include frequency, amplitude, and waveform [26]. improves reliability, utilizing diodes for unidirectional current However, the waveform cannot be efficiently exploited for and capacitors as analog memory to retain the peak value. To

4

R ensure subsequent circuit stages remain unaffected, resistor Vfout Vdout Vfout R2 facilitates efficient charge discharge, isolating downstream C stages, necessitating component values where 1/(2πR2 C2 ) ≈ ′ ′ Vdout fin matches the attacker-injected frequency fin . (a) Charge accumulation circuit. Reference voltages determined by R3 and R4 necessitate sufCharge Charge Vout Accumulation Release ficiently large R4 /(R3 + R4 ) ratios to reject noise, exemplified by our R4 · Vcc /(R3 + R4 ) > 0.2 V requirement countering Vdout sub-0.2V noise. Simulation results in Fig. 5(a) and Fig. 5(b) alongside parameter settings in Fig. 19 demonstrate peak (b) Latch circuit. (c) The waveforms in charge accumulation circuit. detector output amplitude scaling with input frequency. Finally, a comparator isolates signals exceeding threshold Fig. 6: The duration detector schematics and the simulation voltages, mitigating filter imperfections to prevent noise- waveforms. induced false activation. We utilize a 1 kΩ pull-up resistor R5 to enhance load handling capability. The parameter requirement a prescribed temporal order. In both modes, circuits consisting for triggering is: of diodes, capacitors, and resistors retain each detection result s for a configurable duration. If a required feature is missing R12 R4 Vin (1) or arrives outside its allowed window, activation is blocked; − Vd > Vcc 1 2 R + R R1 + ω 2 C 2 3 4 1 during sequential detection, the process returns to its initial where Vin is the input signal and Vd is the diode’s forward stage. The payload is activated only after the complete feature voltage. set has been verified with the required timing. 3) Duration Detector: The duration detector can be realized Amplitude. To ascertain whether a measured voltage falls within a specific threshold range, attackers can employ various using a linear integrator circuit, a ramp generator, or similar circuits such as average detectors, limiters, and comparators. We topologies. We adapt the concept of counter-based triggers utilize a window comparator circuit to exemplify the amplitude prevalent in digital circuits [17], by leveraging charge accudetection mechanism, which consists of two comparators as mulation in capacitors coupled with a latch to achieve similar depicted in Fig. 4(b). This topology guarantees high input effects, as depicted in Fig. 6. We omit duration detectors for resistance and produces a high level activation signal exclu- extra transducer Trojans to minimize component counts, since sively when the voltage falls within the designated bounds. The manufacturers and users commonly restrict testing for their comparator configuration relies on the amplitude relationship intended sensing modalities in common practice. Charge accumulation circuit. True trigger condition logic between the injected and legitimate signals. For instance, an amplitude threshold configuration is required when the injected initiates capacitor charging, defining trigger duration as the signal exceeds the legitimate signal’s normal operational range. voltage rise interval to threshold Vth . The minimum voltage of logic high serves as the threshold voltage, eliminating the The window comparator output is defined as: ( need for extra comparators. Consequently, injection duration V cc if Vin ∈ [V− , V+ ] Vf out = (2) Tc can be calculated as: 0 if Vin ∈ / [V− , V+ ] Tc = RC(ln Vf out − ln(Vf out − Vth )) (3) where Vf out is the comparator output, Vin is the sensor voltage, and V− , V+ represent the lower and upper reference voltages, where Vf out is the desired output of the trigger condition respectively. Balancing power consumption and interference detector and R is the series resistance. immunity mandates mega-ohm range level selection for R1 , In practical implementations, capacitors exhibit natural R2 , and R3 , while pull-up resistor R4 stabilizes the operational leakage due to material imperfections and ambient conditions. amplifier output by enhancing load handling and preventing We leverage this inherent discharge mechanism to reset the state undefined floating states. without additional components, preventing charge accumulation AC signal amplitude detection requires diode rectification from compromising the trigger duration detector’s accuracy. and filtering to yield smooth DC signals, then processed via Physical setup and validation results are presented in Fig. 6. DC trigger logic to ascertain input amplitudes. Unless otherwise specified, we utilize R = 470 kΩ and 2) Multi-Condition Detector: To improve robustness and C = 10 µF , resulting in a 4-second rise time, which is suppress false triggers, the detector requires multiple features sufficient to evade factory testing yet enable efficient attacks. of the same input signal, such as amplitude and frequency, Latch. When exploitable signal nodes are limited to sensor to satisfy predefined conditions. In the combinational mode, outputs, activation can disrupt the triggering condition. We feature detectors operate in parallel, and the trigger is asserted therefore employ latch mechanisms to preserve the activated only when all detector outputs are valid within a specified time state independently of subsequent signal variations. Persistent window. In the sequential mode, analog switches control the effects demand strict logic low initialization achieved via a activation of successive detectors: each detector is enabled only capacitor-based power-on-reset design, where capacitors ensure after the preceding feature has been detected, thereby enforcing reliable initialization with negligible power overhead.

5

4) Payload: Payload implementation can utilize CMOS line, we propose asymmetric cutoff techniques introducing DC transmission gates, BJT switches, or MEMS relays to mod- offsets when injected AC signals exceed dynamic range. ify sensor outputs. We employ analog switches to toggle Three methods induce asymmetric clipping. First, modifying transmission paths and mimic broken or shorted states as DC bias resistors shifts the static operating point from VCC /2, required to force the sensor’s output to switch to a preset altering the positive and negative range to cause asymmetric voltage signal for static spoofing. Alternatively, the latch-up saturation. Second, selecting amplifiers with lower Vmax mechanism induces direct supply-to-ground shorts, leading restricts output limits despite identical power supplies. Third, to PDoS. In devices lacking protection mechanisms, the inserting resistors between the main VCC and the amplifier resulting overcurrent causes permanent physical damage, such power port reduces the supply voltage to Vcc′ above normal as trace burnout or component degradation, rendering the sensor output voltages to maintain operation. This induces sensor permanently inoperable. In contrast, although modern voltage drops, lowering saturation thresholds without affecting devices equipped with voltage regulators often incorporate normal operation. thermal shutdown or current limiting, the Trojan can exploit With the injected AC signal included in Vin , the amplifier these safeguards by persistently re-triggering the fault. This output is clipped and exhibits a DC offset. Filtered amplitude forces the sensor into a continuous shutdown loop, indirectly Vout follows: Z t+Td but effectively achieving a PDoS condition. Finally, dynamic 1 spoofing mandates extra transducers to modulate sensor output, Vout (t) = Vaout (τ ) dτ. (4) Td t enabling real-time control of injected values to realize complex where Td denotes the injection signal duration and Vaout (τ ) time-varying waveforms. is the amplifier output. Theoretical maximum offsets (Vmax − ′ Vmax )/2 enable AM signals to spoof any waveform within this Insight 1: LRD Trojans trade structural stealth for range when the resonant period T satisfies T ≪ Td . Saturation controllability, using explicit trigger and payload circuits Trojans require legitimate signal amplitudes below injection to enable reliable triggering and diverse attack effects levels to ensure proper sensor operation. Simulations with across sensor types. the parameters in Appendix E2 yield results in Fig. 7(a), ′ confirming that input amplitudes exceeding Vmax undergo C. Physical-Effect-Driven Trojan Design clipping, inducing output deviations. While LRD Trojans enable diverse attacks, they prove Rectification. The nonlinear rectification effect in amplifiers complex and resource-intensive. Leveraging internal physical can convert high-frequency AC signals into DC offsets. This effects, we design stealth-oriented mechanisms with lower phenomenon stems from the inherent nonlinearity of the hardware costs. Avoiding explicit trigger logic, our approach input differential pair’s current-voltage characteristics. However, emulates analog circuit characteristics via subtle parame- the susceptibility to this rectification varies markedly across ter modifications or minimal additions, eliminating explicit different amplifier topologies. Consequently, adversaries can activation circuits to enhance Trojan stealthiness. Utilizing strategically exploit this variability by targeting amplifiers with components within standard manufacturing tolerances causes pronounced nonlinear sensitivity to maximize the efficacy of forensic analysts to classify Trojans as manufacturing defects physical signal injection attacks. The resulting attack impact or cost-cutting measures rather than malicious components, is illustrated in Fig. 7(b). making them harder to detect during standard inspection. Envelope detection. Beyond enhancing existing nonlinearity, Injected signals target frequency bands beyond LPF cutoff inducing new nonlinearity enables Trojan insertion. For nonfrequencies, such as resonant frequencies, yet face attenuation linear components, a diode is preferred first as it is commonly from sensor filters and protective circuitry. To bypass defenses, used in sensors to block reverse current without disrupting our Trojan translates high-frequency signals into low-frequency normal circuit behavior. The diode’s forward conduction does ones while preserving payload information, detailed in Fig. 3. not impact the circuit’s operation, as it works with the DC bias Unlike LRD Trojans, this conversion typically relies on the configuration of operational amplifiers. sensor’s existing coupling path rather than an additional To enable targeted frequency signal demodulation, we trigger circuit. We therefore systematically analyze each sensor design a circuit comprising a diode, capacitors, and resistors, component and identify exploitable analog effects that can be allowing Trojan insertion via RC parameter tuning for envelope induced or amplified with minimal hardware changes. detection. Furthermore, absent physical diodes, amplifiers 1) Amplifier: We propose four methods for inserting Trojans enforce unidirectional flow mimicking diode behavior to extract into amplifiers: saturation, envelope detection, rectification, and baseband signals, albeit within restricted amplitude range. over-amplification, corresponding to enhancing nonlinearity, Carrier frequency dictates injection efficiency, requiring a inducing nonlinearity, and using a nonlinearity-free approach. balance among sensor frequency response and baseband signal Saturation. Saturation-based Trojans exploit amplifier ampli- bandwidth, while determining demodulation parameters and tude limits for baseband signal recovery. We analyze saturation attack effectiveness. Diode conduction during positive highphenomena in Appendix D to identify exploitable methods. frequency cycles rapidly charges capacitors via low forward Since symmetrical saturation clipping preserves the DC base- resistance. Signals dropping below the capacitor voltage trigger

6

Vcc

V’cc

+ Amp. -

R1

Vin

Vsat

Vout

Amp.

R2

Deviation

C

Vee

Vin 1k

(a) The saturation circuit and the simulation waveforms.

Saturation circuit

Vin

10k

+

R

Vin C

Vout Trojan

Vout NoTrojan

Vout NoTrojan

Vout Trojan

Amp.

Vout

(12.57, 97.28)

_

R1

R2

0.5

C

(c) The envelope detection circuit and the simulation waveforms.

Envelope detection circuit

Vin129MHz

(b) The rectification circuit and theSimulation tested waveforms. Rectification circuit waveforms

Simulation waveforms

Vout

Amp.

Vin262MHz

Vout

(d) The over-amplification circuit and the simulation waveforms.

Over-amplification circuit

Simulation waveform

Simulation waveforms

Fig. 7: The circuit schematics and experimental results of the PED Trojans in the amplifier.

reverse bias, causing cutoff and slow discharge through load 2) Transducer: The transduction process is inherently nonresistors. Brief conduction combined with discharge durations linear and can be formulated as: exceeding signal periods enables AM waveform envelope 2 k Vout (t) = A1 Vin (t) + A2 Vin (t) + ... + Ak Vin (t) (6) approximation. Diodes mimic open circuits for low-frequency legitimate signals, preserving quality. for the input signal and Ak is the gain Simulations involving circuits with and without diodes where A1 is the gain k of the kth term V . in confirm successful baseband signal restoration for both Trojan Nonlinearity commonly arises from manufacturing imperfecmethods, utilizing schematics in Fig. 7(c) and parameters in Appendix E3. Modifying RC-tuned envelope detectors or tions and intrinsic material properties. Components designed adding diodes enables hidden signal demodulation from high- for the same function can vary significantly in their degree frequency carriers without disrupting normal operation. Enve- of nonlinearity, especially in the A2 coefficient. Attackers can lope detection implementation necessitates sensor tolerance for exploit this variation by selecting components with a higher A2 small voltage drops, as added diodes introduce slight voltage to insert a Trojan. To activate the Trojan, amplitude modulation is used to restore the baseband signal via nonlinearity. More reductions. Over-amplification. Beyond nonlinearity, attackers induce sensitive transducers can further enhance the attack efficiency. 3) LPF and ADC: Ideal LPFs block all signals above overamplification via frequency response modification, using high-frequency triggers to drive sensors into saturation and their cutoff frequency. However, real filters only attenuate cause DoS. As temperature and pressure sensors operate in low- these signals. Attackers can increase the cutoff frequency by frequency bands, we ensure stealth by selecting triggers outside modifying the resistance or capacitance. This change reduces this legitimate bandwidth. Inserting a capacitor C parallel to the attenuation of a high-frequency attack signal, allowing it to the input resistor R1 leverages capacitive bypassing at high pass through the filter. To prevent aliasing, the ADC’s sampling rate must be at least twice the LPF’s cutoff frequency according frequencies, yielding gain as: to the Nyquist criterion. The attack works by injecting a signal Vout R2 whose frequency is higher than half the sampling rate. This = (5) 1 Vin R1 ∥ jωC aliasing causes the ADC to misinterpret the high-frequency attack signal as legitimate, low-frequency sensor data. where Vin and Vout denote signals and R1 , R2 , and C values 4) Payload Design: PED Trojans can be designed without corresponding to the circuit components. a distinct payload module. The trigger signal is crafted to be Signal passage at the preset frequency sharply reduces the demodulated or amplified by the circuit itself, which in turn combined C and R1 impedance, increasing gain. We determine executes a DoS or spoofing attack. This approach presents a C via the desired amplification factor, simulating the circuit in trade-off: gaining stealth by eliminating a payload component Fig. 7(d) with parameters in Appendix E4. Results show 12.57 at the cost of reduced attack generality. kHz inputs cause operational amplifier amplification nearly 200 times that of legitimate signals, creating interference, while Insight 2: PED Trojans tune native analog effects to low-frequency inputs remain accurate. Adding a small resistor induce or amplify exploitable vulnerabilities, improving mitigates high-frequency environmental noise triggering to structural stealth without explicit trigger logic. further enhance stealth.

7

TABLE I: Comparison of direct sensor attacks and Trojans. Laptop Laptop

Sensor

Obj.

Direct

Trojan

Impact

Range = 0.5 m -2.2∼ +5.6 °C

Range = 5.1 m -11.6∼ +27.1 °C Valid Command 159◦ Offset Preset Constant Preset Constant Preset Constant Preset Constant

Range +920%

Oscilloscope Oscilloscope Oscilloscope Oscilloscope Signal Generator Signal Generator

Microphone

1

Temperature

1

Microphone

2

Inclination

2

Noise Correct Reading Linear Output Correct Reading Correct Reading Correct Reading

Signal Generator Signal Generator ADCADC Power Amplifier Power Amplifier Ultrasonic Ultrasonic Speaker Speaker

4.9× Gain Command Injection

Power Amplifier Power Amplifier Sensor Sensor

(a) Low-frequency Laptop signal emission.

Near-field Near-field Probe Probe victim's victim's device device

(b) EMI signal emission. Oscilloscope

Oscilloscope Signal Generator Fig. 8: Experimental setup showing attacker devices and Signal Generator Near-field victim sensors. The attacker’s infrastructure consists of a signal ADC Power Amplifier Probe Power Amplifier generator, a power amplifier, and signal emitters. victim's Ultrasonic

∆ = 159◦

Logic device Speaker Sensor Inversion Logarithmic Near-field False Laser Logarithmic Near-field Ultrasonic Speaker Sound Speaker Laser Ultrasonic Speaker Sound Speaker Infrared 3 Antenna Probe Antenna Probe Alarm Control Pressure 3 Loss Loop Logarithmic Near-field Pointer Ultrasonic Speakers Sound Speaker Temperature 3 Antenna Probe Hijack Obj. : Objective; 1: Efficacy Enhancement; 2: Feasibility Expansion; 3: Fig. 9: Signal emitters used in our evaluation across different Functional Extension. attack modalities. Light

3

Key results summary. Our evaluation demonstrates significant improvements over prior sensor attacks. A comprehensive comparison is summarized in Table I and detailed below: • Light sensors: Unlike traditional laser attacks that mainly amplify sensor readings, our Trojans can force persistently low outputs, potentially disrupting greenhouse lighting control and reducing crop quality [33]–[35]. • Microphones: We extended the DoS range from 0.5 meters to 5.1 meters and achieved a 100% spoofing success rate, unlike prior work yielding only noise under identical methods, enabling the manipulation of vehicular voice assistants to unlock doors or control smart home systems. • Infrared sensors: We spoof infrared sensors previously unexplored to enable suppression and false alarm attacks against intrusion detection systems, achieving 100% similarity without unintended triggers. • Pressure sensors: We spoof brake pressure sensors previously unexplored to induce critical braking failure, causing the simulated vehicle to collide with obstacles. • Inclination sensors: We induced sensor-level angle deviations of up to 159°, while the benign sensor remained stable under the same attack setup. In STK simulations, a 30° Trojan-induced attitude deviation caused a right roll and subsequent altitude loss. • Temperature sensors: We amplified temperature deviations by 4.9 times compared to physical signal injection alone, extending the range from negative 11.6 °C to positive 27.1 °C, potentially causing chemical reactor runaway and industrial boiler explosions.

V. E VALUATION To comprehensively evaluate SenTrojan, we conduct extensive real-world experiments across multiple sensor types and attack scenarios. We begin with an overview of our experimental methodology and key findings in Sec. V-A, followed by detailed experimental setups and evaluation metrics in Sec. V-B. We then analyze the Trojan efficacy and hardware overhead of SenTrojan. A. Overview We evaluate a total of 14 sensors across 6 categories to demonstrate the broad applicability of SenTrojan, as detailed in Table II. Our evaluation reveals consistently high successful trigger rates (82%-100%), excellent stealthiness (similarity >97%), and zero false trigger rates for LRD Trojans across all tested sensor types. All similarity errors are below 5%, which is comparable to the maximum specified measurement error of representative commercial sensors [27]–[32], which indicates that the Trojan-induced deviations are unlikely to be distinguishable through routine functional testing. Furthermore, we evaluate common hardware-overhead metrics to quantify the performance impact of SenTrojan. Results show that Trojaninserted sensors exhibit less than 5% deviation from benign sensors in time drift, noise floor, and response time, indicating minimal impact on normal sensor operation. Our Trojan design process begins by examining the transducer type and determining whether the target sensor can receive remotely injected out-of-band signals. If not, an additional transducer can be introduced. We then characterize the injected signal feature, i.e., amplitude, frequency, or their combination, and implement a feature extraction circuit for reliable triggering. We modified one representative sensor for each case, and all implemented designs are included in the reported results.

B. Evaluation Setup 1) Hardware Configuration: The experimental setup comprises four components: trigger signal generation, attack devices, Trojan implementation, and sensor measurement.

8

10.0mm

(a) SPV0842LR5H microphone.

10.0mm

10.0mm

10.0mm

23.3mm 23.3mm

Center 22.3 Max >160.0 Min 18.6

V

I

T

A

L

overheat

close 25.0mm 25.0mm

the

door

overcurrent

(b) S150B381 microphone. （a）

Fig. 10: PED Trojans embedded in two microphone types using PCB. Red boxes indicate the inserted Trojan circuits.

（b）

（c）

Fig. 11: (a) Thermal imaging capture of overheating eventually leading to PDoS, alongside the parameter display of the power supply; (b) Static spoofing results for the microphone; (c) Dynamic spoofing results for the inclination sensor.

Attack device configuration. For optical, ultrasonic, and acoustic attacks, an arbitrary waveform generator [36] combined with a power amplifier [37] drives transmitters to emit number of benign trials conducted without the predefined precise physical signals, using an oscilloscope for real-time trigger signal. For PED cases, a false activation is identified monitoring as depicted in Fig. 8(a). Similarly, EMI attacks when the output deviation from the benign reference exceeds employ an RF signal generator [38] and amplifier [39] to drive 5%, corresponding to the maximum specified measurement a logarithmic antenna or near field probe, shown in Fig. 8(b). error among representative commercial sensors [27]–[32]. Trigger signal generation. Addressing real-world deploy- • Successful trigger rate (STR) is the ratio of trials in ment, we target remotely deliverable optical, acoustic, and which the Trojan is successfully activated and produces electromagnetic triggers using the emitters displayed in Fig. 9. the intended effect to the total number of trials conducted Trojan implementation strategy. For experimental flexwith the predefined trigger signal. We evaluate each Trojan ibility, we implement SenTrojan using commercial off-theover 50 independent trigger trials and calculate STR. shelf sensors and prototype the added Trojan circuitry on Since the different effects of LRD Trojans share a common circuit boards. These prototypes are intended to validate attack triggering mechanism and differ only slightly in payload functionality rather than represent the final physical form of a circuitry, the evaluation of attack effectiveness and stealthiness supply-chain implant. In a practical attack, the inserted circuitry is described together. A PED Trojan primarily enables spoofing can be miniaturized using standard PCB and surface-mount attacks with minimal circuit modifications. design techniques. To demonstrate feasibility, we further build Details for hardware overhead metrics are as follows: a PCB-integrated PED Trojan for a microphone, where the • Time drift describes how sensor output changes over time. complete Trojan-enhanced module measures 10 mm × 25 mm, • Noise floor denotes intensity of inherent background noise. and the Trojan itself occupies approximately 6 mm × 2 mm, • Response time shows how fast the sensor reacts to changes. as shown in Fig. 10. Accordingly, we focus our evaluation on • Power represents the electrical power required by the sensor Trojan efficacy and hardware overhead rather than physical during operation. miniaturization. For hardware overhead, we report the normalized deviation 2) Evaluation Metrics: We evaluate the Trojan efficacy from the benign sensor, where lower values indicate smaller of SenTrojan using three primary metrics and the hardware overhead: overhead using four metrics, with detailed metric information n 1 X m̃i − mi as follows: Overhead m = (8) n i=1 mi • Similarity assesses how Trojan affects the sensor’s normal function by comparing the output of the benign sensor with where mi and m̃i denote the benign and Trojan-inserted that of the Trojan-inserted one. It is calculated as: measurements for metric m, respectively. n X For each Trojan type, we first introduce the applied sensor 1 x̃i − xi Similarity = 1 − (7) and detail the corresponding test procedure used to evaluate n i=1 xi Trojan efficacy on a case-by-case basis. We then present the where n is the total number of the data points, xi is the hardware-overhead evaluation in a unified manner. value from the benign sensor, x̃i is the value from the Trojaninserted sensor. For complex signals like audio waveforms, C. Amplitude Mechanism of LRD Trojan we apply frequency-domain transformations and use spectral 1) Ambient Light Sensor: Ambient light sensors are widely distance metrics. In each case, similarity is calculated over all used in smart agriculture and industrial combustion systems to sampled points rather than from a single measurement. For measure light intensity. A successful static spoofing attack is example, the light sensor evaluation collects approximately defined as the intensity dropping below 10 lux (simulating 108,000 samples over three 10-minute trials under different darkroom or flameout conditions), potentially leading to lighting conditions. disrupted crop growth cycles and hazardous fuel accumulation. • False trigger rate (FTR) is the ratio of benign trials in We design amplitude detection triggers detailed in Appendix F1. which the Trojan is unintentionally activated to the total We executed 50 trials at 1 m across five diurnal periods,

9

TABLE II: Overall performance of SenTrojan.

Type

SenTrojan SSI+ Add. Cond. Mod. DC Sensor

Type

Manuf.

Trojan Efficacy Model

Hardware Overhead

Sim.↑ FTR↓ STR↑ Drift↓ Noise↓ RT↓ Power↓

Adafruit Vishay Light 3+8 Siemens Amphenol TRUE Temp. Heraeus CUI Dev. FALSE Mic. DFRobot Ultras. 3+12 Murata Infrared TRUE Adafruit Pressure Dayang

GL5506 TEMT6000 QRB1 NHQ103B PT100 CMC-9745 Sen0487 IRA-S210 SR501 DYTB-002

99.7% 99.9% 99.4% 99.0% 97.8% 99.7% 99.9% 100% 100% 99.3%

0% 0% 0% 0% 0% 0% 0% 0% 0% 0%

100% 82% 98% 100% 96% 96% 98% 94% 96% 96%

0.5% 0.3% 0.5% 0.9% 1.7% 1.4% 0.2% 0.9% 0.1% 0.3%

0.3% 0.5% 0.9% 0.9% 0.7% 0.8% 0.2% 0.7% 0.6% 0.7%

0.2% 0.5% 0.8% 4.4% 3.4% 1.4% 2.4% 1.5% 0.9% 0.3%

0.8% 7.4% 6.6% 13.6% 0.5% 10.2% 8.7% 0.6% 6.7% 3.9%

ED

Ultras.

0+1

FALSE

Mic.

Satu.

Sound

0+1

FALSE

Inclin.

Transd.

EMI

0+0

FALSE

Temp.

Sen0487 ADXL335 ADXL337 RETCB

99.8% 99.4% 99.2% 99.1%

0% 0% 0% 0%

100% 100% 98% 96%

0.2% 0.5% 0.8% 1.1%

0.7% 0.5% 0.9% 0.5%

0.8% 3.5% 4.4% 3.2%

7.8% 16.5% 15.9% 0.6%

Amp.

LRD

Freq.

PED

Sensor

FALSE

Light

DFRobot Analog Analog People

Mod.: Modality; SSI: Small Scale Integration; DC: Discrete Component; Sim.: Similarity; FTR: False Trigger Rate; STR: Successful Trigger Rate; RT: Response Time; Amp.: Amplitude; Freq.: Frequency; ED: Envelope Detection; Satu.: Saturation; Transd.: Transducer; Cond.: Condition; Add.: Additional; Manuf.: Manufacturer; Mic.: Microphone; Temp.: Temperature; Ultras.: Ultrasound; Inclin.: Inclination; CUI Dev.: CUI Devices.

yielding 100%, 82%, and 98% success rates as the laser overcame natural ambient fluctuations. The STR of TEMT6000 is lower than that of other LRD cases because its transducer package attenuates the injected laser and reduces the effective injection strength. Continuous 72-hour monitoring yielded zero false triggers, while three 10-minute comparative trials under identical lighting revealed average similarities of 99.67%, 99.88%, and 99.35%, confirming negligible operational impact. In addition, the Trojan latch mechanism requires a sensor reset to recover from persistent spoofing.

disabling the sensor even when irreversible physical destruction is prevented. 2) Temperature Sensor: In addition, we employ this Trojan mechanism on a temperature sensor, incorporating a photodiode as an extra transducer. This Trojan executes static spoofing by clamping the sensor reading to a falsified low temperature, thereby forcing the thermal control system into a continuous heating state, potentially causing overheating. The trigger success rate was assessed following the identical methodology used for the light sensor. For stealthiness evaluation, continuous sampling was conducted from 20 °C to 80 °C on a controlled heating platform. The results showed no false triggers. Specifically, for NHQ103B and PT100, the Trojan-inserted versions exhibited data similarities of 98.95% and 97.79%, respectively, compared to their benign baselines.

For PDoS attacks, we employed power supply devices equipped with protective features, including overvoltage protection, overcurrent protection, overheat protection, overload protection, and reverse polarity protection, to ensure testing conditions closely resembled real-world scenarios while avoiding potential risks [40]. The results are illustrated in Fig. 11(a). D. Frequency Mechanism of LRD Trojan During PDoS attacks, the Trojan-induced short circuit drew 0.99A and raised the local temperature above 170°C. After 1) Microphone: Leveraging rare high-intensity ultrasound removing the trigger and power cycling, the sensor did not during normal microphone operation, we design frequency recover normal functionality. Crucially, this thermal breakdown detection triggers detailed in Appendix F2. SenTrojan can targets the Trojan’s own inserted analog switch. Because this silence microphone output to 0 V, as shown in Fig. 11(b), engineered short operates outside the sensor’s native current yielding 96% and 98% success rates across 50 trials at a path, localized overcurrent protections are inherently bypassed. 20 cm distance for two different sensor models. Maximum Unless strict global current limiters are enforced at the main attack distance evaluations using identical ultrasound injections power supply, native safeguards fail to prevent this localized reveal that the Trojan-inserted Sen0487 achieved a 5.1 m DoS destruction. Furthermore, even if modern industrial devices are range versus 0.5 m for standalone attacks at equal power, equipped with robust current-limiting regulators and thermal whereas the CMC-9745 was immune to DoS attacks without the protections, the Trojan effectively exploits these very safeguards. Trojan. Stealthiness tests simulating household environments By intentionally driving the device into a sustained state of with MUSAN audio [41] of 10 hours produced no false thermal or overcurrent stress to trigger continuous protection triggers. Similarity between benign and Trojan-inserted sensors cycles, the Trojan achieves a persistent DoS attack, successfully for Sen0487 and CMC-9745 exhibited 99.9% and 99.7%,

10

PC

Trojan-inserted Sensor Infrared Detector

Inserted Trojan

Benign Sensor Results

Benign Simulation PC

No one detected

Human detected

Speaker

Input

Power Line

Victim Sensor

Trojan-inserted Sensor Results Human detected

No one detected

Attack Simulation

Experiment setup

Vifa Speaker

Simulation scenarios

Simulation results

Fig. 13: Evaluation of a pressure brake sensor with a HIL driving simulation.

Fig. 12: Infrared intrusion-detection experiment. The red indicator denotes human-presence detection; after Trojan activation, the detector reports an incorrect state.

Attack Result

1.56

Amplitude (V)

Output

Hidden Trojan

1.54 1.52 1.50 open

the

door

respectively, across 2 hours of recording via normalized log 0.2 0.3 0.4 0.5 Time (s) spectral distance [42], confirming minimal operational impact. Fig. 14: The waveform of Fig. 15: The recognition re2) Trojan Application: Integrating an extra microphone the injected command to sult of the Trojan-based attransducer enables ultrasonic triggering for other sensors. spoof the microphone. tack on three ASR models. Infrared sensor. Infrared sensors detect human presence via a digital output generated from an infrared-sensitive resistor and a comparator. The infrared sensor is installed at an office doorway and monitors over a continuous 4-hour period along with the pre-collision speed and travel distance profiles, to evaluate the false trigger rate and similarity. The sensor is shown in Fig. 13. accurately identified all 47 individuals with zero false positives, maintaining 100% average similarity to the benign baseline E. The PED Trojan of Microphone under normal operation. In addition, we employ this Trojan in a PED Trojans utilize amplitude-modulated voice embedded commercial intrusion detection system that uses infrared sensors in carriers to spoof sensors. During Trojan activation, they to detect intruders and trigger alarms. At a distance of 1 meter, demodulate high-frequency ultrasound to recover baseband the attacker used a Vifa speaker [43] to emit the ultrasonic commands like opening the door, potentially triggering unautrigger signal. Under this setup, the Trojan-inserted sensor thorized actions as shown in Fig. 14. Our setup employs could be induced to either miss a human-presence event and computer-generated voice basebands and 27 kHz carriers to output “No one detected,” or falsely report a human presence, evaluate effectiveness following NormDetect protocols [46], whereas the system reported correct detection under normal [47]. Testing recognition across IFlytek [48], OpenAI [49], operation, as shown in Fig. 12. These effects enable attackers to and Azure [50] ASR systems yielded results in Fig. 15. evade infrared detection and potentially conduct unauthorized To mitigate nonlinearity, we first injected 27 kHz electrical activities such as trespassing or accessing restricted areas. signals directly into amplifier ports, achieving 100% success. Pressure sensor. Linearly indicating brake magnitude, the Second, leveraging prior ultrasonic methods [4], we reduced pedal pressure sensor constitutes a critical component for safe sampling rates to 16 kHz, sufficient for speech. As ultrasonic driving. We develop a Hardware-In-the-Loop (HIL) simulation components alias through the ADC to eliminate nonlinear utilizing CarSim [44] and MATLAB Simulink [45] to illustrate impacts, we distinguished Trojan-induced behavior from system the control-level consequence of the Trojan in a closed-loop functionality, demonstrating 100% success. Conversely, benign simulation environment. In this setup, the pressure sensor output microphones yielded only noise under identical injection due is sampled by an Arduino Uno and forwarded to the PC-based to weak demodulation. Finally, we assessed similarity and vehicle controller, allowing us to compare benign and Trojan- STR using the same PED Trojan evaluation protocol. The triggered conditions under the same control logic and obstacle microphone achieved a 100% STR over 50 trigger trials. scenario. For the benign condition, the vehicle initially travels By placing the device separately in laboratory and office at a constant speed, and braking is applied when an obstacle environments for ten hours, we observed no false ASR sentence is detected ahead. With the benign sensor input, the controller activations (FTR = 0%). receives a valid brake signal, the vehicle speed decreases to zero Application. Modern cars are equipped with voice virtual as expected, and the car stops before reaching the obstacle. assistants that use ASR models to interpret user commands. To For the attack condition, once the Trojan is triggered, the demonstrate the threat of microphone spoofing, we conduct an compromised pressure sensor output is forced to 0, causing the experiment where commands demodulated by the Trojan are controller to miss the braking event. Under this controlled HIL injected into the victim’s microphone in the car to closely configuration, the simulated vehicle therefore fails to decelerate simulate real attack scenarios. The results reveal that the in time and collides with the obstacle. The experimental setup, injected command causes the car door to open, potentially

11

leading to dangerous situations, such as unauthorized access to the vehicle or theft.

PC

Benign Simulation

F. The PED Trojan of Inclination Sensor Attack Simulation

Speaker Extensively deployed in drones, smartphones, and automotive Victim safety systems [51], inclination sensors measure acceleration Sensor and angles. We investigate PED Trojan designs targeting Experiment setup Simulation scenarios Simulation results modules dependent on signal processing circuits. Frequency sweeps identify z-axis resonance at 720 Hz for the ADXL335 Fig. 16: Evaluation of the inclination sensor in the flight control and 518 Hz for the ADXL337. Controlled by trigger intensity system. When the Trojan is triggered, the aircraft undergoes and amplifier asymmetry, DC amplitude manipulation is shown an attitude tilt and its altitude decreases rapidly. in Fig. 11(c). SenTrojan induces 159° and 26° biases for the tested sensor models, which are the maximum sensorlevel deviations measured in our PED Trojan experiments. Victim 44.5℃ 44.5℃ Victim Thermostat Because PED Trojans lack the well-defined activation threshold Thermostat 30.4℃ 30.4℃ characteristic of LRD Trojans, we first establish a reference 17.4℃ 17.4℃ output in a preliminary test. We then define that the PED 17.4℃ 17.4℃ 11.4℃ 11.4℃ 5.8℃ Trojan is successfully activated if the result of a subsequent test Antenna Antenna 5.8℃ Modified Wire Modified Wire falls within ±20% of this reference value. Conversely, benign sensors exposed to identical triggers exhibit no distortion due (a) Setup. (b) Result. setup result to insufficient asymmetry for saturation. Finally, comparing setup result temp_values = [17.4, 30.4, 44.5, 17.4, 11.4, 5.8] benign and Trojan-inserted sensors across nine angles yielded Fig. 17: The temperature-controlled temp_values = system [17.4, 30.4, 44.5,wire 17.4, length 11.4, 5.8] with 99.43% and 99.23% similarity, confirming low observability adjustment serving as the Trojan mechanism. under normal operation. We further evaluated the FTR across nine angles, repeating each measurement five times and collecting 10 s of data at each angle. The results showed transducer wire-length modification, which is common in output similarities of 99.4% and 99.2% for the ADXL335 and chemical reactors and industrial manufacturing [53], [54]. ADXL337, respectively, corresponding to deviations below 5%, Modifying transducer wire lengths enables frequencywith an FTR of zero for both sensors. Over 50 trigger trials, dependent attacks. Specifically, 10 cm wires yield 11.6°C the ADXL335 and ADXL337 achieved STRs of 100% and drops at 789 MHz and 27.1°C increases at 710 MHz, while 1 98%, respectively. cm wires exhibit 2.2 °C drops and 5.6 °C increases at 1185 Application. Widely applied in robotics and aviation, in- MHz and 1298 MHz, respectively, as illustrated in Fig. 17. clination sensors can directly affect flight stability and safety, This enhancement amplifies the variation by 4.9× without and constitute potential factors in crashes [1]. We compute hardware component changes. Furthermore, adjusting signal the position coordinates and body-attitude quaternions from amplitude controls deviations, enabling attackers to trigger the sensor outputs and import them into STK [52] for flight thermal runaway, production hazards, or financial loss. We simulation under matched benign and attack conditions. We further evaluated the FTR across ten temperatures, repeating feed ADXL335 measurements into the STK simulation and each measurement five times. The results showed an output introduce a Trojan-induced 30° attitude deviation, which is similarity of 99.1%, corresponding to a deviation below 5%, sufficient to disrupt attitude control and initiate a right roll. with an FTR of zero. Over 50 trigger trials, the RETCB Under the benign condition, the aircraft maintains a stable temperature sensor achieved an STR of 96%. attitude and altitude. Under the attack condition, the erroneous sensor measurement causes the flight controller to issue H. Hardware Overhead To evaluate the hardware overhead of SenTrojan, we opposing corrective commands, producing a right roll. We further incorporate a lift-loss model to capture the effect of compare the factory-calibrated performance of benign and the increasing roll angle on altitude. As shown in Fig. 16, the Trojan-inserted sensors under identical environmental and physical stimuli. For each metric, we report the normalized attack induces an attitude tilt and rapid altitude loss. deviation measured under the same test condition. The resulting G. The PED Trojan of Temperature Sensor deviations are summarized in Table II, where lower values Temperature sensors are critical in medical and industrial sys- indicate smaller functional perturbation and better stealth. tems, where erroneous readings can disrupt closed-loop thermal Time drift. We collect 600 samples at 10 Hz immediately control and contribute to unsafe operating conditions, poten- after initialization and another 600 samples after 24 hours of tially including excess cooling scenarios such as the British dye continuous operation, and compare the corresponding sensor factory incident [3]. Targeting temperature-controlled outlets outputs under identical conditions. The comparison shows no that regulate power based on real-time thermal feedback, long-term DC offset beyond the baseline variation, with the we examine a PED Trojan mechanism implemented through measured normalized deviation remaining below 2% across all

12

tested sensors. This result is consistent with the high-impedance design of the LRD monitoring circuit, which reduces loading on the native signal path. The added parasitic capacitance is also small relative to the circuit capacitance, limiting its impact on long-term output stability. (a) EM shielding. (b) Sound shielding. Noise floor. We quantify the noise floor using the average peak-to-peak voltage measured under identical test conditions. Fig. 18: The shielding effects of different materials. The measured noise-floor deviation remains below 1% across all tested sensors. For LRD Trojans, the comparator inputs operate in a high-impedance state, which limits leakage-induced B. Active Detection current noise. For PED Trojans, the low-frequency powerActive protection employs EM leakage analysis and funcsupply rejection of the analog front end attenuates noise tional testing to distinguish benign from Trojan circuits, introduced by the added series components, limiting its effect utilizing runtime anomaly detection to identify triggers. within the measured operating bandwidth. Active detection based on EM leakage. Detection proves Response time. We evaluate the response time using the difficult as our attack exploits (1) triggers activating signals 10% to 90% step-response rise time. The measured normalized beyond normal ranges, (2) high-impedance parallel or shortdeviation ranges from 0.2% to 4.4% across the tested sensors. circuit series inactivity modes minimizing power and circuit For LRD Trojans, the added series analog switch contributes performance impact, and (3) subtle parametric modifications only Ron ≈ 6 Ω, and the parallel monitoring circuits do obscured by component density. EM-based detection methods not modify the original sensor signal path, resulting in little exploit circuit-level modifications introduced by sensor Trojans, change to the output RC time constant. For PED Trojans, the which alter the resulting electromagnetic leakage profiles. inserted series resistors and diodes have limited impact on Clustering algorithms then use these leakage patterns to transient response. The resistors keep transient currents within distinguish benign circuits from Trojan-infected ones. However, the operational amplifier’s driving capability, the diodes add this approach is ineffective against PED Trojans because their little dynamic resistance during current spikes, and the analog small component-parameter perturbations produce negligible front end compensates for small supply-voltage drops. These changes in EM leakage. design choices reduce Trojan-induced circuit overhead and Detection based on functional testing. Extending standard match the measured rise-time results. evaluations [25], functional testing exposes physical triggers Power. We evaluate power overhead using a DC power by varying amplitudes, frequencies, and modalities, including supply to monitor sensor power consumption. For LRD Trojans, acoustic, optical, and electromagnetic. Conversely, duration- or the normalized power overhead remains below 10% on average. pattern-dependent activations evade detection due to impractical This result is consistent with the use of LinCMOS components requirements for long-duration injections. Impedance-based tamper detection. For pre-deployment and mega-ohm resistors, which reduce power consumption and keep the Trojan power consumption at the microwatt verification, an LRD Trojan employs high-impedance parallel level. For PED Trojans, the power change depends on the monitoring. Its added components may alter the overall circuit insertion mechanism. Inserted diodes can slightly reduce overall impedance, enabling detection using methods [55] when a power consumption by lowering the downstream load voltage, trusted reference profile is available. A PED Trojan miniwhereas some series-resistor PEDs produce a normalized power mizes component modifications and therefore causes smaller deviation above 10% in low-power sensors. Nevertheless, sensor impedance perturbations. For example, a modification that power consumption can vary substantially under different changes only the wire length may be difficult to detect. For physical stimuli during normal operation because different runtime impedance-based monitoring, an activated LRD Trojan stimulus intensities induce different current levels. For example, toggles the signal path, thereby changing the circuit impedance. a microphone can consume more than twice the power This change can potentially expose the Trojan to impedance measured under no-sound input, showing that the observed monitoring against a trusted reference. Runtime anomaly detection. Complementing prePED overhead is comparable to normal stimulus-dependent application detection, runtime verification targets correlation power variation. errors and spectral noise anomalies, mitigating prohibitive VI. P OTENTIAL D EFENSES physical isolation costs. Utilizing sensor fusion for consistency We propose passive shielding and active detection to defend checks while scanning transducer outputs for distortion from common injections, including out-of-band frequencies, this against physical signal injection in SenTrojan. monitor effectively identifies attack footprints. A. Passive Shielding VII. R ELATED W ORK As the primary defense against EM, acoustic, and optical signals, passive shielding uses metal casings to block EM A. Sensor Attacks coupling, opaque materials to prevent lasers, and soundPrior research has revealed that a wide variety of sensor absorbing foam to dampen acoustic waves, as shown in Fig. 18. types across multiple fields are vulnerable to security threats.

13

For example, modulated lasers can inject voice commands into microphones [56], while electromagnetic interference can induce false touch events on capacitive touchscreens [57]. These threats have been demonstrated in automotive systems [58]– [61], drones [24], [62], smart devices [4], energy systems [63], medical devices [5], and other applications [64]. This research often involves introducing attack signals through physical media, such as light [56], acoustic waves [4], electromagnetic waves [5], or electric fields [65]. Despite extensive research uncovering diverse sensor attack vectors, existing approaches remain fundamentally constrained by their reliance on preexisting hardware vulnerabilities and continuous external signal emission. As a result, their effectiveness is often sensor-specific, and their impact is difficult to sustain once the attack signal is removed. In contrast, SenTrojan investigates internal hardware compromises via sensor Trojans, which embed malicious functionality into sensor circuits to make previously resistant sensors susceptible to crafted physical triggers. This shift from exploiting to implanting vulnerabilities broadens the sensor attack surface, enabling more generalizable, controllable, and persistent attacks. B. Hardware Trojan Attacks Hardware Trojans, typically realized through malicious circuit modifications [13], have been widely studied in IC security, particularly for high-stakes military systems. Prior work examined untrusted military ICs [66] and kill-switch threats [67], both involving digital circuitry insertion. Compared with IC-level Trojans, board-level security remains less explored [68], motivating automated synthetic Trojan insertion into PCB designs [69]. Existing Trojans are commonly classified as logic-based designs [7], [8], which add gates, or dopant-based designs [70], [71], which modify physical properties. While Yang et al. [12] replace digital counters with analog capacitors for Trojan triggering, their design still monitors internal digital states and targets digital logic. In contrast, SenTrojan targets analog sensors, where crafted external physical signals activate hidden analog behavior without software or digital access. Comparisons of SenTrojan with A2 [12] are shown in Appendix G. This shift introduces analog-specific challenges, namely preserving benign sensing accuracy while ensuring reliable activation under environmental noise and component variations. Moreover, by leveraging native physical effects such as transducer nonlinearity and amplifier saturation, SenTrojan reduces reliance on explicit logic anomalies and can evade conventional digital-integrity checks and visual inspections. VIII. C ONCLUSION In this paper, we introduce SenTrojan, a novel threat vector manipulating sensor outputs via malicious hardware modifications. We propose two distinct trigger mechanisms: Logic-Rule-Driven (LRD) Trojans relying on dedicated blocks for specific signal patterns, and Physical-Effect-Driven (PED) Trojans exploiting inherent behaviors. Extensive evaluations across 14 sensors confirm reliable triggering with minimal

14

operational disruption. Finally, we explore countermeasures mitigating impacts, emphasizing designer vigilance regarding this emerging threat. R EFERENCES [1] Curt Devine and Drew Griffin. Boeing relied on single sensor for 737 max that had been flagged 216 times to faa, 2019. https://edition.cnn.co m/2019/04/30/politics/boeing-sensor-737-max-faa/index.html. [2] Jack Stewart. Why tesla’s autopilot can’t see a stopped firetruck. 2018. [3] Process Safety Beacon. Excess cooling can cause a runaway reaction, 2018. https://www.aiche.org/resources/publications/cep/2018/july/process -safety-beacon-excess-cooling-can-cause-runaway-reaction. [4] Guoming Zhang, Chen Yan, Xiaoyu Ji, Tianchen Zhang, Taimin Zhang, and Wenyuan Xu. Dolphinattack: Inaudible voice commands. In Proceedings of the 2017 ACM SIGSAC conference on computer and communications security, pages 103–117, 2017. [5] Denis Foo Kune, John Backes, Shane S Clark, Daniel Kramer, Matthew Reynolds, Kevin Fu, Yongdae Kim, and Wenyuan Xu. Ghost talk: Mitigating emi signal injection attacks against analog sensors. In 2013 IEEE Symposium on Security and Privacy, pages 145–159. IEEE, 2013. [6] Zhicong Zheng, Jinghui Wu, Shilin Xiao, Yanze Ren, Chen Yan, Xiaoyu Ji, and Wenyuan Xu. Phyfuzz: Detecting sensor vulnerabilities with physical signal fuzzing. In NDSS, 2026. [7] Swarup Bhunia, Michael S Hsiao, Mainak Banga, and Seetharam Narasimhan. Hardware trojan attacks: Threat analysis and countermeasures. Proceedings of the IEEE, 102(8):1229–1247, 2014. [8] Kan Xiao, Domenic Forte, Yier Jin, Ramesh Karri, Swarup Bhunia, and Mohammad Tehranipoor. Hardware trojans: Lessons learned after one decade of research. ACM Transactions on Design Automation of Electronic Systems (TODAES), 22(1):1–23, 2016. [9] Endres Puschner, Thorben Moos, Steffen Becker, Christian Kison, Amir Moradi, and Christof Paar. Red team vs. blue team: a real-world hardware trojan detection case study across four modern cmos technology generations. In 2023 IEEE Symposium on Security and Privacy (SP), pages 56–74. IEEE, 2023. [10] Komail Dharsee and John Criswell. Jinn: hijacking safe programs with trojans. In 32nd USENIX Security Symposium (USENIX Security 23), pages 6965–6982, 2023. [11] Yiming Li, Yong Jiang, Zhifeng Li, and Shu-Tao Xia. Backdoor learning: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2022. [12] Kaiyuan Yang, Matthew Hicks, Qing Dong, Todd Austin, and Dennis Sylvester. A2: Analog malicious hardware. In Proceedings of the 2016 IEEE Symposium on Security and Privacy (SP), pages 18–37. IEEE, 2016. [13] Mohammad Tehranipoor and Farinaz Koushanfar. A survey of hardware trojan taxonomy and detection. IEEE Design & Test of Computers, 27(1):10–25, 2010. [14] W Göpel, José Hesse, and JN Zemel. Sensors. A Comprehensive Survey. Vol.: 1 Fundamentals and General Aspects. VCH, 1989. [15] John F Miller. Supply chain attack framework and attack patterns. The MITRE Corporation, MacLean, VA, 2013. [16] Jie Zhang, Feng Yuan, and Qiang Xu. Detrust: Defeating hardware trust verification with stealthy implicitly-triggered hardware trojans. In Proceedings of the 2014 ACM SIGSAC conference on computer and communications security, pages 153–166, 2014. [17] Adam Waksman and Simha Sethumadhavan. Silencing hardware backdoors. In 2011 IEEE Symposium on Security and Privacy, pages 49–63. IEEE, 2011. [18] Timothy Trippel, Kang G Shin, Kevin B Bush, and Matthew Hicks. Bomberman: Defining and defeating hardware ticking timebombs at design-time. In 2021 IEEE Symposium on Security and Privacy (SP), pages 970–986. IEEE, 2021. [19] Frank Imeson, Saeed Nejati, Siddharth Garg, and Mahesh Tripunitara. Non-deterministic timers for hardware trojan activation (or how a little randomness can go the wrong way). In 10th USENIX Workshop on Offensive Technologies (WOOT 16), 2016. [20] Asmit De, Mohammad Nasim Imtiaz Khan, Karthikeyan Nagarajan, and Swaroop Ghosh. Hartbleed: Using hardware trojans for data leakage exploits. IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 28(4):968–979, 2020.

[21] Omer Shwartz, Amir Cohen, Asaf Shabtai, and Yossi Oren. Shattered trust: When replacement smartphone components attack. In 11th USENIX Workshop on Offensive Technologies (WOOT 17), 2017. [22] MITRE Corporation. Supply Chain Compromise: Compromise Hardware Supply Chain, 2023. https://attack.mitre.org/techniques/T1195/003/. [23] Information Commissioner’s Office. Supply Chain Attacks: Learning from the Mistakes of Others. https://ico.org.uk/about-the-ico/research-r eports-impact-and-evaluation/research-and-reports/learning-from-the -mistakes-of-others-a-retrospective-review/supply-chain-attacks. [24] Timothy Trippel, Ofir Weisse, Wenyuan Xu, Peter Honeyman, and Kevin Fu. Walnut: Waging doubt on the integrity of mems accelerometers with acoustic injection attacks. In 2017 IEEE European symposium on security and privacy (EuroS&P), pages 3–18. IEEE, 2017. [25] Frank E McFadden and Richard D Arnold. Supply chain risk mitigation for it electronics. In 2010 IEEE International Conference on Technologies for Homeland Security (HST), pages 49–55. IEEE, 2010. [26] Shilin Xiao, Wenjun Zhu, Yan Jiang, Kai Wang, Peiwang Wang, Chen Yan, Xiaoyu Ji, and Wenyuan Xu. Sok: Understanding the fundamentals and implications of sensor out-of-band vulnerabilities. arXiv preprint arXiv:2508.16133, 2025. [27] NXP Semiconductors. MPX5010 Integrated Silicon Piezoresistive Pressure Sensor, 2024. https://www.nxp.com/docs/en/data- sheet /MPX5010.pdf. [28] Amphenol NovaSensor. NPC-300 Disposable Medical Pressure Sensor. https://amphenol-sensors.com/en/product-spotlights/npc-300-disposabl e-medical-pressure-sensor. [29] Honeywell. Midas S2 Electrochemical Hydrogen Sensor, 2024. https: //prod-edam.honeywell.com/content/dam/honeywell-edam/sps/hgas/ja/ products/high-tech-and-semiconductor/midass2/documents/hon-ia-pmc -midas-s2-e-h2x.pdf. [30] Amphenol Advanced Sensors. A-1737 Motor Coil Temperature Sensor, 2017. https://amphenol-sensors.com/hubfs/Documents/AAS-920-540 D-Thermometrics-A-1737-MotorCoil-121917-web.pdf. [31] Analog Devices, Inc. ADXL321 Accelerometer, 2007. https://www.anal og.com/media/en/technical-documentation/data-sheets/ADXL321.pdf. [32] Infineon Technologies AG. IM72D128V01 MEMS Microphone, 2022. https://www.infineon.com/assets/row/public/documents/24/49/infineon-i m72d128-datasheet-en.pdf?fileId=8ac78c8c80027ecd0180612a60277e 10. [33] Erik Runkle. Grower 101: Controlling photoperiod, 2002. https://www. canr.msu.edu/resources/grower 101 controlling photoperiod. [34] Thomas Boyle. Photoperiod control systems for greenhouse crops, 2017. https://www.umass.edu/agriculture-food-environment/greenhouse-flori culture/fact-sheets/photoperiod-control-systems-for-greenhouse-crops. [35] University of Wisconsin–Madison Division of Extension. Supplemental lighting, 2026. https://fyi.extension.wisc.edu/energy/greenhouses/supple mental-lighting/. [36] SIGLENT. SDG6032X, 2021. https://siglentna.com/product/sdg6032x/. [37] Micronix. NF HSA4051, 2013. https://eshop.micronix.eu/measuremen t-equipment/electrical-quantities/nf-corporation-instruments/high-speed -bipolar-amplifiers/hsa-4051.html. [38] Rigol. DSG3136b, 2019. https://rigolshop.eu/rf-signal-generator-dsg31 36b. [39] Mini-Circuits. HPA-272+. https://www.minicircuits.com/WebStore/dash board.html?model=HPA-25W-272%2B. [40] Alientek. DP100 Digital Power Supply, 2024. https://en.alientek.com/Pr oduct Details/45.html. [41] David Snyder, Guoguo Chen, and Daniel Povey. Musan: A music, speech, and noise corpus. arXiv preprint arXiv:1510.08484, 2015. [42] Augustine Gray and John Markel. Distance measures for speech processing. IEEE Transactions on Acoustics, Speech, and Signal Processing, 24(5):380–391, 1976. [43] Avisoft Bioacoustics. Ultrasonic dynamic speaker vifa, n.d. https: //avisoft.com/playback/vifa/. [44] Mechanical Simulation Corporation. Carsim, 2023. https://www.carsim .com/products/carsim/index.php. [45] MathWorks, Inc. Hardware-in-the-loop, 2023. https://ww2.mathworks. cn/en/discovery/hardware-in-the-loop-hil.html. [46] Xinfeng Li, Xiaoyu Ji, Chen Yan, Chaohao Li, Yichen Li, Zhenning Zhang, and Wenyuan Xu. Learning normality is enough: a software-based mitigation against inaudible voice attacks. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2455–2472, 2023.

15

[47] Fengchen Yang, Wenze Cui, Xinfeng Li, Chen Yan, Xiaoyu Ji, and Wenyuan Xu. Lightantenna: Characterizing the limits of fluorescent lamp-induced electromagnetic interference. In NDSS, 2025. [48] iFLYTEK. iFLYTEK ASR. https://global.xfyun.cn/products/speech-to-t ext. [49] OpenAI. OpenAI ASR. https://platform.openai.com/docs/guides/speech -to-text. [50] Microsoft. Microsoft Azure ASR. https://azure.microsoft.com/en-us/pr oducts/ai-services/ai-speech/. [51] Christopher J Fisher. Using an accelerometer for inclination sensing. AN-1057, Application note, Analog Devices, pages 1–8, 2010. [52] Ansys Government Initiatives. Systems tool kit (stk), 2023. https: //www.ansys.com/products/missions/ansys-stk. [53] Instrumart. Temperature controller basics handbook, 2018. https://www. instrumart.com/pages/283/temperature-controller-basics-handbook. [54] West Control Solutions. Temperature monitors and limiters, 2018. https: //www.west-cs.com/assets/Brochures/BR-LD-2-US-1903-web-1.0.pdf. [55] Tahoura Mosavirik, Fatemeh Ganji, Patrick Schaumont, and Shahin Tajik. Scatterverif: Verification of electronic boards using reflection response of power distribution network. ACM Journal on Emerging Technologies in Computing Systems (JETC), 18(4):1–24, 2022. [56] Takeshi Sugawara, Benjamin Cyr, Sara Rampazzi, Daniel Genkin, and Kevin Fu. Light commands: Laser-based audio injection attacks on voicecontrollable systems. In 29th USENIX Security Symposium (USENIX Security 20), pages 2631–2648, 2020. [57] Yan Jiang, Xiaoyu Ji, Kai Wang, Chen Yan, Richard Mitev, Ahmad-Reza Sadeghi, and Wenyuan Xu. Wight: Wired ghost touch attack on capacitive touchscreens. In 2022 IEEE Symposium on Security and Privacy (SP), pages 984–1001. IEEE, 2022. [58] Hocheol Shin, Dohyun Kim, Yujin Kwon, and Yongdae Kim. Illusion and dazzle: Adversarial optical channel exploits against lidars for automotive applications. In Cryptographic Hardware and Embedded Systems–CHES 2017: 19th International Conference, Taipei, Taiwan, September 25-28, 2017, Proceedings, pages 445–467. Springer, 2017. [59] Jonathan Petit, Bas Stottelaar, Michael Feiri, and Frank Kargl. Remote attacks on automated vehicles sensors: Experiments on camera and LiDAR. Black Hat Europe, 2015. White paper. [60] Ruchir Chauhan. A platform for false data injection in frequency modulated continuous wave radar. Utah State University, 2014. [61] Wenyuan Xu, Chen Yan, Weibin Jia, Xiaoyu Ji, and Jianhao Liu. Analyzing and enhancing the security of ultrasonic sensors for autonomous vehicles. IEEE Internet of Things Journal, 5(6):5015–5029, 2018. [62] Yunmok Son, Hocheol Shin, Dongkwan Kim, Youngseok Park, Juhwan Noh, Kibum Choi, Jungwoo Choi, and Yongdae Kim. Rocking drones with intentional sound noise on gyroscopic sensors. In 24th USENIX security symposium (USENIX Security 15), pages 881–896, 2015. [63] Jayaprakash Selvaraj, Gökçen Yılmaz Dayanıklı, Neelam Prabhu Gaunkar, David Ware, Ryan M Gerdes, and Mani Mina. Electromagnetic induction attacks against embedded systems. In Proceedings of the 2018 on Asia Conference on Computer and Communications Security, pages 499–510, 2018. [64] Mordechai Guri, Boris Zadov, and Yuval Elovici. Led-it-go: Leaking (a lot of) data from air-gapped computers via the (small) hard drive led. In Detection of Intrusions and Malware, and Vulnerability Assessment: 14th International Conference, DIMVA 2017, Bonn, Germany, July 6-7, 2017, Proceedings 14, pages 161–184. Springer, 2017. [65] Seita Maruyama, Satohiro Wakabayashi, and Tatsuya Mori. Tap’n ghost: A compilation of novel attack techniques against smartphone touchscreens. In 2019 IEEE Symposium on Security and Privacy (SP), pages 620–637. IEEE, 2019. [66] Dean R Collins. Trust, a proposed plan for trusted integrated circuits. 2006. [67] Sally Adee. The hunt for the kill switch. IEEE Spectrum, 45(5):34–39, 2008. [68] Huifeng Zhu, Xiaolong Guo, Yier Jin, and Xuan Zhang. Pcbench: Benchmarking of board-level hardware attacks and trojans. In Proceedings of the 26th Asia and South Pacific Design Automation Conference, pages 396–401, 2021. [69] Aritra Bhattacharyay, Shuo Yang, Jonathan Cruz, Prabuddha Chakraborty, Swarup Bhunia, and Tamzidul Hoque. An automated framework for board-level trojan benchmarking. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 42(2):397–410, 2023. [70] Georg T Becker, Francesco Regazzoni, Christof Paar, and Wayne P Burleson. Stealthy dopant-level hardware trojans. In International

Conference on Cryptographic Hardware and Embedded Systems, pages 197–214. Springer, 2013. [71] Wei Hu, Chip-Hong Chang, Anirban Sengupta, Swarup Bhunia, Ryan Kastner, and Hai Li. An overview of hardware security and trust: Threats, countermeasures, and design tools. IEEE Transactions on ComputerAided Design of Integrated Circuits and Systems, 40(6):1010–1038, 2021. [72] Marco Tartagni. Electronic sensor design principles. Cambridge University Press, 2022. [73] John G Webster and Halit Eren. Measurement, Instrumentation, and Sensors Handbook: Two-Volume Set. CRC Press, 2018. [74] Cadence. Pspice. https://www.cadence.com/en US/home/tools/pcb-desig n-and-analysis/analog-mixed-signal-simulation/pspice.

A PPENDIX A. Life Cycle of Sensor Sensor life cycles typically span five phases: design, fabrication, distribution, maintenance, and recycling [7], [26], [72], [73]. (1) Design: Engineers define structures, component parameters, and types for chips and peripheral circuits based on application requirements. (2) Manufacturing: Following design finalization, factories produce chips and board-level circuitry. While final testing ensures standards, cost constraints often limit manufacturers to basic functionality checks, omitting thorough safety evaluations [25]. (3) Distribution: Supply chains deliver standalone chips or fully assembled modules for system integration. (4) Maintenance: High-end sensors require regular calibration, diagnostics, and component replacement to ensure consistent long-term performance. (5) Recycling: Endof-life recovery and refurbishment support cost savings and environmental sustainability. B. The Adversary’s Capabilities of Different Stages Design stage. One attack scenario arises when malicious actors infiltrate the sensor design team and obtain whitebox access to circuit parameters and component choices, a threat model widely explored in prior work [16]–[19]. Specifically, malicious insiders covertly modify hardware components to embed Trojans before delivery. This stage is feasible because sensor design, verification, and subsequent fabrication involve different personnel or organizations. However, the modification must evade design review, calibration, and functional validation, while preserving sensing range, noise, drift, power, and normal behavior. Routine black-box tests may not exercise a hidden trigger, but an independent golden model may reveal unauthorized changes. However, the golden model is unavailable when the same party controls both the modification and the reference artifacts. • Manufacturing stage. Another attack scenario occurs during sensor fabrication [10], [12], [20], when an untrusted foundry manufactures the sensing chip. Adversaries with access to fabrication masks or process parameters can introduce layout- or process-level modifications as Trojans without changing the overall circuit architecture. This stage is feasible because fabrication is often separated from sensor design and may be performed by an untrusted party. However, analog modifications must remain effective despite process variation while preserving reliability and normal sensing behavior. Moreover, a trusted design file may enable physical or electrical comparison and expose discrepancies. •

16

•

Post-manufacturing stage. Attackers may also insert Trojans after manufacturing is complete [21]–[23], including during PCB assembly and maintenance. In modern sensor systems, discrete sensing elements, signal-conditioning circuits, and sometimes microcontrollers are integrated on the PCB. Attackers involved in assembly or maintenance may replace legitimate components with compromised ones, insert additional circuitry, or modify PCB traces to manipulate sensor outputs. This stage is feasible because assembly, repair, calibration, and maintenance provide direct access to board-level components and interconnects. However, the attacker must overcome physical-access constraints, enclosure restrictions, limited board area, and the need to preserve normal operation after rework. Because the sensor has typically completed factory-side hardware inspection and validation by this stage, a Trojan that preserves normal functionality and avoids unintended activation is difficult to identify through routine post-service testing.

C. Trigger Signal Modality We evaluate ultrasound, laser, EMI, and sound signals across five critical dimensions: flexibility, distance, effect, generality, and difficulty. • Trojan flexibility. Signal modality dictates injection nature and design constraints. Laser offers superior flexibility, injecting fixed-intensity and modulated signals. Ultrasound and sound rank secondary, limited to modulated signals, while EMI exhibits the lowest flexibility, as high-frequency signals induce capacitive and inductive failures. • Attack distance. Laser achieves maximum range via minimal attenuation, followed by directional ultrasound. Conversely, sound propagates omnidirectionally, while EMI relies on restricted near-field coupling. • Attack effect. Trigger modalities influence outcomes. Laser, ultrasound, and sound support broad attack spectra, including DoS and electrical faults. Conversely, EMI is limited to spoofing as high-frequency behaviors alter component responses, reducing versatility. • Attack generality. EMI demonstrates the highest generality as most sensors possess coupling accessible transmission lines. In contrast, other modalities rely on specific transducer types, limiting injection efficiency and applicability. • Attack difficulty. Laser attacks present the greatest challenges due to visibility and targeting precision. While audible and detectable, sound simplifies deployment via omnidirectional propagation. Finally, ultrasound and EMI offer maximum stealth, remaining inaudible with relaxed targeting requirements. D. Analysis of Saturation The output range of an amplifier is determined by the value of the power supply and the circuit structure, where the maximum output value is Vmax , and the minimum value is Vmin . Saturation is common in analog circuits and refers to a quantity that cannot exceed either Vmax or Vmin . The quiescent point is the DC bias Vq required for an amplifier to

make the transistor operate in the linear region. It is necessary For example, the range of an iron thermocouple is -210°C to for the proper operation of the amplifier. There is no fixed 1200°C, meaning a possible spoofing attack of ±141◦ C can standard value for the bias voltage in single-supply amplifiers, be introduced. which depends on the specific type and application scenario. VCC The voltage corresponding to the DC quiescent point is usually 1k 2.5V set to (Vmax + Vmin )/2 to provide the maximum dynamic Vin Vout Vsat range. The output signal of the amplifier is: 330 LM358 100k  LM358 VOFF=0  if Ad Vin + Vq ≤ Vmin Vmin 10μ 10μ VAMPL=0.1 Vaout = Ad Vin + Vq if Vmin ≤ Ad Vin + Vq ≤ Vmax FREQ=1k  AC=0.5  Vmax if Ad Vin + Vq ≥ Vmax 20k 1μ 1k (9) where Ad is the differential gain of the input voltage and Vin Fig. 20: The PSpice simulation circuit of a saturation Trojan is the amplitude of input signal. in the amplifier. When the injected AC signal exceeds the dynamic range, 3) The Envelope Detection Effect Simulation: We simulate the output signal will typically be symmetrically clipped( represented as Vmax − Vq = Vq − Vmin ) by the amplifier the envelope detection amplifier circuit and introduce a diode due to the design of the quiescent point. The AC signal will as the Trojan. The specific circuit is shown in Fig. 21 and Fig. 22. In Fig. 21, the injected signal has a 27 kHz carrier then be removed by the low-pass filter (LPF). modulated by a 500 Hz baseband signal. In Fig. 22, the carrier E. The Circuit Simulation Details and baseband frequencies are 100 Hz and 1 Hz, respectively. The frequencies of these two signals determine the parameTo better understand the function of different components ters of the subsequent demodulation circuit. The demodulation intuitively, we simulate the circuits using the PSpice model [74]. circuit recovers the baseband signal. For low-frequency signals, 1) Frequency Detector Simulation: We simulate the trigger the capacitor, due to its high impedance at low frequencies, mechanism using an LM393 comparator as illustrated in Fig. 19. can be considered an open circuit and thus does not affect the The circuit employs a front-end rectifier to convert the AC input circuit’s performance. into a DC voltage Vp. When the input amplitude generates a Vp exceeding the reference threshold set by the resistor network, VCC VCC the comparator output toggles. This mechanism confirms that Vbb injected AC signals can effectively manipulate the digital output 100k Vin Vout state without affecting normal operation at lower amplitudes. AN4558 VOFF=0.1V VAMPL=0.05V FREQ=500 AC=0.5

VCC Vf1 Vin

VOFF=1 VAMPL=0.5 FREQ AC=1

Vp

1μ

9k

100k

2k

2k

0.5μ

0.5μ

Fig. 21: The PSpice simulation circuit of an envelope detection Trojan in the amplifier.

LM393 330

200

1k

1000n

1k

23k

1μ

VOFF=0 1000n VAMPL=1 FREQ=27k AC=1

20k 2k VOFF=1V VAMPL=0.5V FREQ=1 AC=0.5

Fig. 19: The PSpice simulation circuit of the frequency detector.

VCC

Vin

Vbb VOFF=0V VAMPL=1V FREQ=100 AC=1

Vout 1k

5m 2k 2) The Saturation Effect Simulation: We simulate the satu1k ration amplifier circuit and introduce a 1kΩ resistor between Vcc and the Vcc port of the amplifier as a Trojan. The specific circuit and results are shown in Fig. 20. When the amplitude Fig. 22: The PSpice simulation circuit of an envelope detection of the injected signal exceeds 3V , the output signal is cut off, Trojan in the amplifier without a diode. inducing a DC offset. When the amplitude of the input signal is lower than 3V , the output is unaffected. The stealthiness 4) The Over-amplification Effect Simulation: Simulations is guaranteed because not all sensor operating ranges are full- depicted in Fig. 23 utilize frequency sweeping to determine scale. In our tests with the common operational amplifiers response curves. Results indicate amplification increases 200 LM358 and LM318, the maximum output of the amplifier times between 12 kHz and 13 kHz relative to low-frequency does not reach Vcc. Finally, the output signal of the low-pass transmission, facilitating DoS attacks. Added capacitors funcfilter, representing the induced offset signal, is about 0.5V . tion as open circuits for low-frequency inputs. Unlike ideal

17

models in prior expressions, practical bandwidth limitations suppress high-frequency signals, preventing unintended triggers. Furthermore, in practical applications, the resistance introduced by wires or the addition of a small series resistor with the capacitor mitigates noise-induced false triggering. Vcc

10k

10k

100k

200k

Fig. 23: The PSpice simulation circuit of an over-amplification Trojan in the amplifier. F. The Trojan Test Details 1) Light-triggered Trojan with Amplitude Mechanism: Depicted in Fig. 24, the Trojan circuit comprises a comparator, analog switch, latch, diode, capacitors, and resistors (3 SSI, 8 DC). A window comparator monitors real-time voltage (0.145 V to 0.970 V) mapped from light intensity. Sustaining signals within this range for 2.5 seconds charges the capacitor, toggling the analog switch to alter signal paths. Upon triggering, an electrical fault attack short-circuits critical wiring, potentially causing overheating and damage. Adapting for additional photoresistor transducers involves adding voltage dividers for detection and routing original sensor outputs to the switch, ensuring normal operation. Duration Detector 470 kΩ

1 MΩ

V-

Switch S

1 μF

Vtested

Vtested

10 μF

1 MΩ

Vtested

10 μF 10 kΩ

1 μF

S

Vout

0

G. Comparison with A2 1000k

1 MΩ

470 kΩ

1 MΩ

Fig. 25: Schematics of a frequency-based LRD Trojan in the Sen0487.

100

0.1m

V+ 5.1 MΩ

Switch

0.2 MΩ

LM358

Vcc

Vtested

Duration Detector

Vcc

0.1 μF

1n

Amplitude Detector

Frequency Detector 0.22 nF

33 kΩ

Vout

Vin

transducers entails configuring basic ECM setups and routing analog switch Vtest inputs to original sensor outputs.

Vout

Vcc

47 kΩ

Fig. 24: Schematics of an amplitude-based LRD Trojan in the GL5506. 2) Ultrasound-triggered Trojan with Frequency Mechanism: Minimizing component count relative to band-pass alternatives, a high-pass filter (33k Ω, 0.22 nF, 21, 933 Hz) processes ultrasound signals, depicted in Fig. 25. Peak detectors extract amplitudes while parallel resistors prevent charge buildup. To mitigate 0.3 V leakage from filter imperfections, a comparator sets a 0.6 V threshold to prevent false triggers, augmented by duration detectors. The payload employs analog switches akin to Sec. V-C1. In single transducer sensors lacking signal conditioning, low microphone outputs and diode drops impede peak extraction. Consequently, attackers insert amplifiers and feedback resistors to boost AC amplitudes, enabling Trojan functionality. In addition, utilizing additional microphone

18

SenTrojan differs from A2 [12] along four dimensions: • Trigger source: A2 is activated by internal digital states through a crafted program, whereas SenTrojan is triggered by external physical signals injected through the sensor. • Insertion point: A2 inserts malicious circuitry into unused layout areas, whereas SenTrojan targets sensor-associated analog front-end and PCB-level circuitry. • Payload: A2 modifies a digital privilege flip-flop after activation. SenTrojan manipulates sensor outputs to induce PDoS, static spoofing, or dynamic spoofing. • Validation: A2 uses fabricated malicious processors and software triggers. SenTrojan is evaluated on 14 sensors spanning six transduction principles, with system-level case studies.

Related documents

• Mission-Aware Attestation Envelopes for Time-Critical Autonomous Action: A Hardware-in-the-Loop V2I Study#1218563
• Mission-Aware Attestation Envelopes for Time-Critical Autonomous Action: A Hardware-in-the-Loop V2I Study#1210184
• CVE-2026-99146#1493702
• Harbormaster: Evidence-Gated, Replay-Safe Maritime Anomaly Detection on AWS#1155454
• CVE-2026-88514 — MEDIUM 5.5#1252379