2026 IEEE 33rd International Conference on Electronics, Circuits and Systems (ICECS)
Assessing Runtime Electromagnetic Detection of CPU Hardware Trojans Targeting Kernel Memory Athanasios Moschos˚ , Baki Berkay Yilmaz: , Kevin Valakuzhy˚ , Angelos D. Keromytis˚ ˚ Georgia Institute of Technology, School of Electrical and Computer Engineering, USA : Wayne State University, Department of Electrical and Computer Engineering, USA
arXiv:2609.23186v1 [cs.CR] 19 Sep 2026
© 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
Abstract—Side-channels are a promising approach for hardware trojan detection, as they can passively reveal malicious hardware activity without requiring additional circuitry or destructive analysis of the device under test. This work investigates the use of electromagnetic (EM) emanations for runtime detection of hardware trojan attacks. Using an open-source hardware trojan that implements an arbitrary memory-write primitive, we tamper with the kernel memory of a Linux system running on a RISCV microarchitecture and perform a real-time baseline detection using the processor’s EM emissions. Our results indicate that, under certain conditions, hardware trojans can be detected indirectly through the anomalous software behavior they enable. Index Terms—Hardware Trojans, Signal Processing, Operating Systems, Computer Architecture, Electromagnetic Side-channels
I. I NTRODUCTION Detecting hardware trojans during chip deployment is becoming increasingly important due to the absence of industrywide standards for identifying malicious hardware logic prior to fabrication. This challenge is especially pronounced in CPUs, whose complex operation and diverse microarchitectural activity produce highly variable execution patterns. In this work, we investigate the runtime detection of hardware trojans that target the dynamically allocated kernel memory within CPUs. Kernel memory is particularly important as it hosts security-relevant runtime objects (e.g., task credentials, process descriptors, security policy structures). We show that electromagnetic emissions encode information about the accessed memory region, kernel or user, allowing us to distinguish malicious activity on dynamically allocated kernel memory. Building on this observation, we develop a golden-free runtime detection method that targets such trojan-assisted activity. To the best of our knowledge, this approach to EM-based trojan detection has not been previously explored. Prior work [1, 2] demonstrated that software-level asymmetries can produce distinct electromagnetic signatures. Specifically, Callan et al. [1] showed how electromagnetic measurements can identify instruction-level execution events. Deviation from expected code-execution patterns is also identifiable [2]. Other EM methods, such as backscattering [3], identify hardware logic modifications directly, but require physical access to the chip surface. In practice, this requirement is
979-8-3195-1905-4/26/$31.00 © 2026 IEEE
often difficult to satisfy for deployed CPUs due to standard packaging and cooling components, such as heatsinks. II. TAMPERING W ITH K ERNEL M EMORY Hardware Trojan Design: We use one of the open-source Interrupt-Resilient Trojan designs [4], IRT-1, that features a trigger mechanism implemented within the processor’s register file. The trojan’s payload circuit is designed to bypass the operating-system-enforced isolation between privileged and non-privileged memory regions. Typically, kernel pages have their user-mode bit cleared to prevent user-level accesses. To enable malicious overwrites from user space, the payload circuit masks the user-mode bit of kernel page table entries (PTEs), thereby suppressing memory-access exceptions. Attack Emulation: Using a custom Linux kernel module (LKM), we allocate a 4 MB integer array within the kernel space and expose its base address to the user space, similar to a memory-disclosure primitive. Array allocation employs either kmalloc or vmalloc. A malicious user-space binary can tamper with kernel memory by targeting the disclosed address. To assess the spectral difference between malicious kernelmemory overwrites and benign user-memory overwrites, the binary allocates a separate 4 MB user-space array through malloc and generates random in-bounds indices. At runtime, the binary’s overwrite routine targets either the user- or the kernel-space array using those indices. Hence, both scenarios have the exact same execution flow. Activating the payload only during the overwrite routine bounds the interval of suppressed memory-access exceptions. III. EM-BASED D ETECTION M ETHODOLOGY To assess the sensitivity of EM-based monitoring to trojan activity, we adapt the supervised-learning methodology of [5] to a one-class classification approach. Our method comprises three stages. Initially, during a two-stage training process, EM activity from various CPU operations is characterized to construct a model of benign EM signatures. In the subsequent testing stage, EM signals captured from live CPU operation are processed and compared against this training model in real time to identify anomalous activity.
2026 IEEE 33rd International Conference on Electronics, Circuits and Systems (ICECS)
Training Stage-1: We acquire long EM waveforms, which we partition into shorter time segments, each with a duration equal to the sample time. Using a preconfigured window size and overlap ratio, we compute the Short-Time Fourier Transform (STFT) for each segment. We combine the resulting sequence of time-frequency representations to construct a twodimensional array analogous to a spectrogram, with one axis being the frequency bins and the other the time. Next, we apply max-pooling to this time-frequency array, motivated by two considerations. First, we expect trojan activity to primarily affect dominant frequencies of the acquired signal. The monitoring system is constantly capturing emanations, unaware of the exact trojan activation interval. Hence, max-pooling preserves the effects of short-lived anomalies appearing in these dominant frequencies. Second, the spectral content at certain frequencies can be unstable, exhibiting smearing across neighboring frequencies. Such behavior could mislead the distance-based classifier of the testing stage, due to the accumulation of small differences across the non-dominant frequency bands. Consequently, max-pooling provides a more compact and robust spectral representation, through a lowerdimensional time-frequency array that preserves anomalous activity in dominant frequencies and reduces sensitivity to local smearing-induced spectral fluctuations. Another consideration is noise present in the measurement setup. To mitigate noise, we average the rows of the reduceddimensionality array, producing a single spectral representation. We convert this representation to dB to allow the distance-based classifier to also capture the relative changes occurring in weaker spectral regions. Fluctuations in the global power level across measurements of the same CPU operation can increase variability. We account for this by applying a normalization scheme: we subtract from each single spectral representation its mean and subsequently divide it by its maximum magnitude. Hence, we preserve the overall spectral shape and minimize the effect of global power fluctuations. Overall, the training stage-1 pipeline is computationally lightweight, and is used both to construct the training-stage EM-signature model and to process the real-time emanations at the testing stage. Consequently, computations in this stage must complete within the waveform acquisition intervals. Training Stage-2: This stage contributes only to the EMsignature model generation. The stage-1 measurements correspond to distinct CPU operations (e.g., memory accessing), each one with multiple spectral representations. Since normalization rescaled every spectral representation to a comparable range, we apply the k-means algorithm to reduce the size of each EM model associated with a CPU operation. The resulting cluster centroids of each EM model compactly represent the dominant CPU activity patterns, forming a unified EMsignature database for efficient inference during testing. Testing Stage: To classify whether live EM signals are the outcome of anomalous activity, we compute the ℓ1 -norm distance between each acquired signal (after stage-1 processing) and the database’s EM-signatures. The smallest resulting distance is then compared against a predefined threshold,
TABLE I: Resource utilization of CVA6 and IRT-1. Module Name
Design LUTs
Design FFs
Design-to-CVA6 LUT Ratio (%)
Design-to-CVA6 FF Ratio (%)
CVA6 w. IRT-1 IRT-1 Trigger IRT-1 Payload
72,371 26 4
46,878 1 3
100 0.0359 0.0055
100 0.0021 0.0064
flagging the signal as anomalous if the threshold is exceeded. IV. E XPERIMENTAL S ETUP Trojan Testbed: Similar to [4], we use the 64-bit CVA6 microarchitecture [6] to implement the IRT-1 trojan. The experiments are performed on the Genesys 2 board, with the trojan-afflicted CVA6 operating at a 50 MHz clock and booting a Linux system based on kernel v5.10.7. The ratio of the malicious logic to the overall size of the trojan-afflicted CVA6 implementation is presented in Table I. Higher ratios yield increased chances of discovering the trojan through its side-channel emanations [7]. In our case, the very small trojan-to-host-design ratio in the generated bitstream challenges our detection approach [8]. Measurement Setup: Our monitoring system uses a TekBox 20 mm near-field magnetic probe to measure EM emanations during CVA6’s operation. The probe is connected to a B205mini-i SDR for signal acquisition, with an Apple M4 MacBook serving as the real-time processing platform. Similar to [3], we measure frequency ranges around the CVA6 clock harmonics. To determine the optimal probe location and center frequency, we manually search for a location beneath the FPGA with strong EM emissions during the execution of memory operations. We find a favorable center frequency at 265 MHz, near the 5th clock harmonic. The SDR’s bandwidth is set to 8 MHz to minimize data throughput and the processing overhead of runtime measurements. Our method is golden-free, hence the trojan-afflicted CVA6 is used to capture measurements during both the training and the testing stage. V. R ESULTS A. Spectral Representation of Memory Overwrites To illustrate the spectral characteristics associated with overwrites in the kernel- and user-space memory, we collect EM signals during five distinct scenarios spanning legitimate and malicious activity. We pass the measurements through the training stage-1 processing before plotting their spectrograms. Kernel Overwrites: We modify the LKM described in Section II to allocate a 4 MB integer array either via kmalloc or vmalloc, and subsequently perform 300 K back-to-back element overwrites at random indices. The spectral characteristics of this procedure are shown in Fig. 1. Notably, overwrites on the vmalloc-allocated object exhibit a significantly stronger EM signature, while taking longer to execute. This temporal and spectral divergence is attributed to differences in the page-table-walking activity. These two memory-allocation functions rely on inherently distinct virtual-to-physical memory-mapping mechanisms. Regions allocated through vmalloc face a substantially higher
2026 IEEE 33rd International Conference on Electronics, Circuits and Systems (ICECS)
Fig. 1: Spectrograms of random kernel overwrites in a 4 MB array allocated via kmalloc (top) and vmalloc (bottom).
Fig. 2: Spectrograms demonstrating a user process performing benign and trojan-assisted memory overwrites. The top shows benign random overwrites on a 4 MB user-space region allocated with malloc. The middle and bottom spectrograms show malicious random overwrites on a 4 MB kernel-space region allocated with vmalloc and kmalloc, respectively.
TABLE II: Hyperparameter configurations per training model. Training Model ID
Size of STFT Window
Overlap Ratio
Max-Pooling Kernel Size
Sample Time (ms)
r1, 2, 3s r4, 5, 6s r7, 8, 9s r10, 11, 12s r13, 14, 15s r16, 17, 18s r19, 20, 21s r22, 23, 24s r25, 26, 27s r28, 29, 30s r31, 32, 33s r34, 35, 36s
1024 1024 4096 4096 1024 1024 4096 4096 1024 1024 4096 4096
0.1 0.5 0.1 0.5 0.1 0.5 0.1 0.5 0.1 0.5 0.1 0.5
4 4 4 4 32 32 32 32 64 64 64 64
r5, 10, 20s r5, 10, 20s r5, 10, 20s r5, 10, 20s r5, 10, 20s r5, 10, 20s r5, 10, 20s r5, 10, 20s r5, 10, 20s r5, 10, 20s r5, 10, 20s r5, 10, 20s
other two. This variation is not the outcome of the malicious hardware switching activity, but rather the absence of the pagetable-walking activity during the malicious overwrites. We must note here that pages larger than 4 KB can also be allocated to user-space processes via the kernel feature of transparent huge pages or THP. This could result in fewer address translations, as in the case of kmalloc. However, THP support for RISC-V was introduced after Linux v5.14, and is unavailable in our kernel version. Moreover, THP behavior is dependent on the kernel configuration and operating system distribution. Thus, it is outside the scope of our evaluation. B. Spectrum-Based Monitoring Evaluation
address-translation overhead, as vmalloc constructs a contiguous virtual region of 4 KB physical pages. For a 4 MBsized array, this results in multiple leaf page table entries. In contrast, kmalloc allocates physically contiguous memory from regions covered by the kernel’s direct mapping of RAM. Since this mapping is typically backed by huge pages [9], an array of equivalent size is covered by only two 2 MB pages, resulting in more efficient memory accesses. User Process Overwrites: For this experiment we use the LKM and the user-space binary described in Section II. We capture measurements while the binary performs benign and malicious overwrites in the allocated arrays. During the overwrite routine, IRT-1 is activated and the exact same code path executes in all cases. The spectrograms for these scenarios can be seen in Fig. 2. The spectral and temporal differences between the top (malloc) and middle (vmalloc) spectrograms are insignificant, as large user-space allocations obtained through malloc are also backed by conventional 4 KB pages. This results in similar demand for address translations and significant page-table-walking activity. In contrast, the bottom spectrogram (kmalloc) presents temporal and spectral discrepancies distinguishable from the
Our training model incorporates the spectral representations shown in Fig. 1 and the user-space overwrite activity of Fig. 2 (top). The malicious activity (middle & bottom) in Fig. 2 is the target behavior that our monitoring system aims to detect during testing and flag it as kernel tampering. The ℓ1 -norm distance metric can reliably distinguish successive overwrites of kmalloc-backed kernel objects, owing to their distinct spectral signature. However, the spectral activity generated by successive overwrites of vmalloc-allocated objects closely resembles the emissions of benign user-space overwrites, making such tampering events hard to detect with our system. In practice, the attack duration is not known a priori. A live monitoring system must therefore detect EM anomalies across attacks of varying duration. We approximate this variability by progressively reducing the number of malicious kernel-space overwrites on kmalloc-allocated memory and evaluating the resulting detection sensitivity across candidate training models. For this quantitative assessment, we use the area under the precision-recall curve (AUPRC) and summarize the results in the heatmap of Fig. 3. We choose AUPRC due to the dataset imbalance between trojan and benign activity. Table II presents the hyperparameters used to construct the candidate models. Models 13–24 consistently perform best across malicious overwrite counts from 100 K down to 10 K, with models 14 and 17 achieving 100% AUPRC even at 10 K; only models 21 and 24 show notably degraded performance at this count. These results indicate that the max-pooling kernel size strongly influences detection performance, with the mediumsized value of 32 yielding the best overall results. Models 32 and 35 also achieve comparable AUPRC at 10 K, while using a larger STFT window size of 4096 samples and a large kernel
2026 IEEE 33rd International Conference on Electronics, Circuits and Systems (ICECS)
Fig. 3: Heatmap evaluating detection capability of distinct training models under varying trojan-assisted memory overwrites.
5K
100
2 1 2 2 1 2 2 0 0 1 0 0 9 7 2 8 10 2 2 0 0 3 0 0 0 0 0 0 0 0 6 14 2 5 16 2
10 K 10 9 6 11 11 6 10 10 12 9 10 12 86 100 94 87 100 89 89 92 28 90 93 24 1 1 0 1 1 0 45 96 30 33 89 33 20 K 14 19 10 15 20 12 15 13 17 13 13 17 92 100 100 88 100 100 84 80 90 85 89 91 0 0 0 0 0 0 66 84 95 52 85 92 50 K 21 43 37 22 44 45 8 8 10 7 8 11 83 97 98 84 95 97 61 77 87 66 89 88 0 0 0 0 0 0 65 71 73 57 72 72
100 K 12 29 26 13 30 30 25 19 24 22 19 24 100 100 100 99 100 100 98 100 100 98 100 100 0 1 0 0 1 0 90 98 100 77 98 100 1
6
11
16
21
Candidate Training Model ID
size of 64. Notably, a 10 ms sample time is common to the high-performing models 14, 17, 32 and 35. Nevertheless, shorter overwrite sequences produce less distinctive spectral patterns, as reflected by the very low AUPRC scores across all models at 5 K overwrites. An attacker who distributes memory modifications across multiple short bursts may therefore evade detection. Future work will focus on improving detection at such low overwrite counts. VI. D ISCUSSION The very small footprint of IRT-1 relative to the overall CVA6, combined with its intertwined placement within the CPU layout, complicates acquisition of a pristine trojan signature at runtime [7, 8]. Spectrally diverse electromagnetic emissions from various microarchitectural components further exacerbate this challenge. Nevertheless, our results indicate that trojan-assisted software execution can produce distinct EM signatures detectable without comparison against trojanfree measurements. We therefore argue that CPU hardware trojans can be detected through the anomalous software behavior they enable. This marks a departure from traditional detection approaches that seek to isolate the minute power or EM variations directly attributable to the trojan logic. To understand whether this shift in detection strategy can practically aid identification of kernel-memory tampering, we examine the prevalence of different kernel memory-allocation mechanisms. We analyzed the most recent Linux kernel source code available at the time of writing (v7.1-rc6) for common allocation APIs, including kmalloc/vmalloc and their variants. Our analysis identified approximately 36ˆ more call sites for kmalloc-style allocators than for vmalloc, with over 36,000 and roughly 1,000 call sites, respectively. These results suggest that untargeted attacks are likely to interact with kernel objects allocated via kmalloc-based mechanisms and therefore exhibit detectable emissions. Conversely, carefully scoped modifications of vmalloc-backed objects may
26
31
36
Area Under the Precision-Recall Curve (%)
Counts of Malicious Kernel Overwrites
AUPRC Performance of Hyperparameter Configurations Across Malicious Overwrite Counts
80 60 40 20 0
evade detection, compromising system security. Consequently, favoring kmalloc-based allocation for security-sensitive kernel objects could improve the runtime detectability of kernelmemory tampering attacks via EM side-channels. R EFERENCES [1] R. Callan, A. Zajic, and M. Prvulovic, “A practical methodology for measuring the side-channel signal available to the attacker for instruction-level events,” in 47th Annual IEEE/ACM International Symposium on Microarchitecture, 2014. [2] A. Nazari, N. Sehatbakhsh et al., “EDDIE: EM-based detection of deviations in program execution,” in Proceedings of the 44th Annual International Symposium on Computer Architecture, 2017. [3] L. N. Nguyen, B. B. Yilmaz et al., “A novel golden-chipfree clustering technique using backscattering side channel for hardware trojan detection,” in IEEE International Symposium on Hardware Oriented Security and Trust (HOST), 2020. [4] A. Moschos, F. Monrose, and A. D. Keromytis, “Towards practical fabrication stage attacks using interrupt-resilient hardware trojans,” in IEEE International Symposium on Hardware Oriented Security and Trust (HOST), 2024. [5] B. B. Yilmaz, E. M. Ugurlu et al., “Detecting cellphone camera status at distance by exploiting electromagnetic emanations,” in IEEE Military Communications Conference (MILCOM), 2019. [6] F. Zaruba and L. Benini, “The Cost of Application-Class Processing: Energy and Performance Analysis of a LinuxReady 1.7-GHz 64-Bit RISC-V Core in 22-nm FDSOI Technology,” IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 2019. [7] L. N. Nguyen, C.-L. Cheng et al., “Creating a backscattering side channel to enable detection of dormant hardware
2026 IEEE 33rd International Conference on Electronics, Circuits and Systems (ICECS)
trojans,” IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 2019. [8] S. Bhunia, M. S. Hsiao et al., “Hardware trojan attacks: Threat analysis and countermeasures,” Proceedings of the IEEE, 2014. [9] J. Corbet. (2022) Solutions for direct-map fragmentation. [Online]. Available: https://lwn.net/Articles/894557/