RF Spectrogram Anomaly Detection with Quantum Kitchen Sinks: Architecture, Representation, and Hardware Validation Abdallah Aaraba∗ , Alexis Vieloszynski† , Remon Polus∗ , Ola Ahmad† , Soumaya Cherkaoui∗ ∗ Department of Computer and Software Engineering, Polytechnique Montréal, Montréal, QC, Canada
arXiv:2607.13897v1 [cs.LG] 15 Jul 2026
Email: [email protected], [email protected], [email protected] † Thales cortAIx Labs, Montréal, QC, Canada Email: {alexis.vieloszynski, ola.ahmad}@thalesgroup.com
examples of these adversarial actions include jamming, sniffing, spoofing, and the injection of anomalous signals [4]. Among these threats, jamming attacks are especially harmful, as adversaries intentionally transmit disruptive signals to interfere with legitimate communications and distort observed spectrum activity [5]. Such interference reduces the signal-to-interferenceplus-noise ratio (SINR), resulting in higher error rates, degraded communication quality, and possible system blockage [6]. Furthermore, the rapid advancement of software-defined radio technologies has significantly lowered the barrier to accessing and deploying jamming devices, making these threats more widespread and accessible [7]. Consequently, both civilian and military wireless systems require robust and reliable spectrum monitoring and anomaly detection mechanisms to ensure secure and resilient operation in the presence of such adversarial activities. Anomaly detection techniques aim to identify both known and unknown interference by detecting deviations from normal signal behavior [8]. Early conventional methods, such as energy detection, cyclostationarity, and matched filtering [9], rely heavily on expert knowledge. Recent works leverage machine learning (ML) on spectrogram datasets for jamming detection [10], using supervised approaches with labeled data [11], [12] and unsupervised methods such as autoencoders and predictionIndex Terms—Quantum machine learning, Quantum Kitchen based models [13], [14], deep Convolutional Neural Networks Sinks, anomaly detection, RF spectrograms, hybrid quantum- (CNN)[15], as well as Generative Adversarial Network (GAN)classical models based frameworks [16]. However, these techniques often suffer from high computational complexity, slow convergence, and I. I NTRODUCTION limited adaptability in dynamic environments. Spectrum sharing is a fundamental concept in cognitive radio Quantum Machine Learning (QML) has emerged at the intersystems, enabling multiple heterogeneous networks to coexist face of quantum computing and classical ML, with the goal of while efficiently utilizing shared radio-frequency (RF) spectrum exploring whether quantum information processing can enhance resources that are inherently limited [1]. This paradigm relies the efficiency or performance of learning models [17]. Building on the open and dynamic nature of the radio spectrum, where on this perspective, QML has also attracted growing interest for wireless transmission propagates over inherently shared chan- anomaly detection in wireless systems, where high-dimensional nels [2]. However, this openness inherently exposes wireless signal representations and complex interference patterns make environments to a wide range of disruptions, including both efficient feature extraction particularly important [18]. In unintentional disturbances and deliberate attacks such as unau- particular, Quantum Kitchen Sinks (QKS) adapt the classical thorized transmissions and malicious interference [3]. Common random kitchen sinks framework of explicit randomized feature Abstract—The broadcast nature of wireless channels exposes radio-frequency (RF) networks to anomalous and malicious transmissions, making anomaly detection a fundamental requirement for secure spectrum management. Quantum Kitchen Sinks (QKS) offer a lightweight hybrid quantum feature map suitable for near-term quantum devices, yet their behavior on structured signal data remains poorly understood. In this paper, we extend the standard QKS template with multi-depth data re-uploading and ring entanglement, and evaluate the resulting pipeline on controlled RF spectrogram anomaly detection. We introduce a validation-locked five-stage ablation protocol that systematically separates the effects of shallow architecture, re-uploading depth, episode budget, input representation, and classical readout. Across the completed benchmark, Discrete Cosine Transform (DCT) representations consistently dominate raw and Principal Component Analysis (PCA) inputs, moderate-depth entangled QKS configurations form the strongest operating regime, and QKS improves over matched classical direct-readout baselines across all evaluated representation-readout pairs on the held-out test set, with the best configuration reaching a test Area Under the Receiver Operating Characteristic curve (AUROC) of 0.8778 and a test F1 of 0.7995. The study bridges two levels of realism: real measured sub-6 GHz cellular signals on the data side and real-device validation on the ibm_quebec Quantum Processing Unit (QPU) on the computing side, with AUROC deviations below 0.013 relative to simulation. These results provide a practical, reproducible framework for deploying QKS-based anomaly detection in wireless networks.
Fig. 1: Diagram of the data acquisition setup. It depicts the measurement of LTE signals, with simulated anomalous signals introduced into the environment. Subsequently, the measured IQ time-domain signals are transformed into spectrograms.
maps to quantum circuits, thereby producing finite-dimensional lightweight classical readout stage. quantum features that can be processed by lightweight classical • We introduce a validation-locked, leakage-free five-stage readouts [19]. QKS is especially attractive for near-term ablation protocol that separates shallow architecture, depth, noisy intermediate-scale quantum (NISQ) devices because the episode budget, input representation, and readout effects quantum circuit acts as a randomized feature generator rather without touching the test split until the final stage. • We show that the Discrete Cosine Transform (DCT) than a fully trainable variational model, while the optimization burden is shifted to a simple classical head [19], [20]. This representations unlock strong QKS performance, with yields a modular hybrid pipeline that is straightforward to multi-depth QKS consistently improving over matched train, compatible with shallow circuit implementations, and baselines across all evaluated readouts on the held-out particularly well suited to controlled ablations across both test set. input representation and circuit-design choices. What remains • Our study bridges two levels of realism by relying on missing in many application-oriented studies, however, is clear real measured data and by validating the pipeline on the characterization of which components of such a hybrid QKS ibm_quebec Quantum Processing Unit (QPU), showing pipeline matter most when the inputs are structured signal deviations in the Area Under the Receiver Operating representations such as RF spectrograms. Characteristic curve (AUROC) below 0.013 relative to simulation. While QML holds significant potential to advance anomaly detection in RF signals, its application to wireless spectrogram The remainder of this paper is organized as follows. Section analysis remains largely unexplored [21]. To the best of II describes the generation of the spectrogram dataset used for our knowledge, QKS has not yet been applied to anomaly both simulations and experiments. Section III introduces the detection in the wireless spectrum, nor validated on real proposed QKS architecture for anomaly classification. Section quantum hardware for this task. Motivated by this gap, this IV presents the performance evaluation of the proposed model. paper presents an end-to-end QKS benchmark for controlled Section V discusses the experimental results. Finally, Section RF spectrogram anomaly detection. We extend the standard VI concludes the paper. shallow QKS template [19] with multi-depth data re-uploading II. S PECTROGRAM DATASET and ring entanglement, and evaluate the resulting pipeline through a validation-locked, leakage-free, five-stage protocol This study leverages the dataset introduced in [22], which that isolates shallow architecture, re-uploading depth, episode contains raw in-phase and quadrature (IQ) samples meabudget, input representation, and classical readout effects in sured directly from real Long-Term Evolution (LTE) transsequence, without touching the test split until the final stage. missions across multiple LTE frequency bands. These realOur study bridges two levels of realism: real measured wireless world measurements serve as the primary user signals in signals on the data side and real quantum hardware execution cognitive radio networks, providing an authentic foundation on the computing side, while anomaly generation remains for evaluating anomaly detection. To simulate adversarial deliberately synthetic and controlled. conditions, anomalous signals are synthetically generated using The contributions of this paper are as follows. Python-based simulations and superimposed onto the measured • We construct a labeled spectrogram dataset combining real LTE data (see Fig.1). The analyzed sub-6 GHz bands are sub-6 GHz cellular signals with synthetically generated actively shared between LTE and 5G New Radio (NR) through anomalous signals spanning three interference types. dynamic spectrum sharing, making the detection of anomalous • We extend the standard QKS template with multi-depth transmissions directly relevant to today’s heterogeneous cellular data re-uploading and ring entanglement, preserving a networks.
(a) Normal Spectrogram
(b) Chirp Anomaly
(c) Barrage Jamming
(d) Frequency Hopping Jamming
Fig. 2: Example RF spectrograms from the dataset used in this study: (a) normal spectrogram; (b)–(d) spectrograms illustrating three different types of anomalies applied to the signal in (a).
In this configuration, each data sample along the time axis corresponds to a duration of 21.15 ms, while the frequency axis spans a bandwidth of 48 MHz, as detailed in Table I. To emulate anomalous behavior within LTE bands, three types of interference signals are synthetically generated and embedded into the dataset. Specifically, the chirp signal represents an unauthorized transmission [23], while the other two—barrage jamming and frequency-hopping noise jamming (a time-varying form of partial-band interference)—correspond to intentional jamming activities [24]. The temporal duration of these signals is randomly selected within 21.15 ms, while their frequency locations are randomly distributed within the 48 MHz band to ensure diversity. The workflow of the dataset is illustrated in Fig. 1. The embedding process is controlled by the Jamming-toSignal Ratio (JSR), which quantifies the relative power of the injected anomalies with respect to the LTE signal. Specifically, JSR values ranging from −10 dB to 5 dB, with increments of 2 dB, are considered to model varying interference conditions. The training set consists of 10,800 unique LTE signals, and the test set consists of a disjoint set of 4,062 signals. Since each signal is included in both its normal form and its anomalyinjected form, this yields the 21,600 training samples and 8,124 test samples reported in Table I.
tinctive signatures associated with chirp, barrage, and frequency hopping patterns, enabling their differentiation from standard LTE activity. The generated spectrograms are subsequently utilized as inputs to the QKS pipeline for anomaly detection. TABLE I: Dataset Summary Train Data samples Test Data samples Sampling Frequency Number of IQ points per Data sample Time Duration Bandwidth Spectrogram Size Anomaly Signal Classes JSR Values [dB]
21,600 8,124 61.44 MHz 1, 300, 000 21.15 ms 48 MHz 400 × 400 Chirp Barrage Jamming Frequency Hopping Jamming [−10, −8, −6, −4, −2, 0, 2, 5]
An example of the generated dataset is illustrated in Fig. 2. The spectrogram of the normal LTE signal, shown in Fig. 2a, exhibits a structured and relatively stable pattern, where the signal energy remains confined within specific By transforming raw IQ samples into spectrograms, both frequency bands corresponding to legitimate transmissions. In legitimate transmissions and anomalous activities can be contrast, the anomalous cases introduce noticeable distortions distinguished through their temporal and spectral characteristics. in the time–frequency representation. As depicted in Fig. 2b, Specifically, a time-domain signal of length 1,300,000 IQ the chirp signal manifests as a continuous diagonal trace, points is formed by combining the original LTE signal with the representing an unauthorized transmission with its time-varying injected anomalies. Subsequently, we generate the spectrograms frequency. In Fig. 2c, the barrage jamming spectrogram by applying a Short-Time Fourier Transform (STFT) using a displays a wideband and dense energy distribution spanning Hann window, with STFT parameters of 3250 and 8192, and the 48 MHz, indicating simultaneous interference across the an overlap ratio of 25% [25]. The resulting time–frequency entire bandwidth. Meanwhile, Fig. 2d illustrates the frequencyrepresentation is subsequently downsampled to a resolution hopping noise jamming, which is characterized by intermittent of 400 × 400. Instead of applying a conventional Fast Fourier bursts that rapidly shift across several subbands, resulting in a Transform (FFT) over the entire one-dimensional IQ sequence, scattered and discontinuous pattern. the STFT partitions the signal into overlapping segments and III. QKS P IPELINE AND A BLATION P ROTOCOL computes the Fourier transform on each windowed portion. These localized spectra are then concatenated along the A. Input representations temporal dimension to form a comprehensive time–frequency Let x ∈ R400×400 denote an RF spectrogram. The QKS representation. This approach facilitates the extraction of dis- pipeline does not operate directly on this matrix, but rather on
(1)
RY (ϕY,1 )
(1)
RY (ϕY,2 )
(1)
RY (ϕY,3 )
q1
RX (ϕX,1 )
q2
RX (ϕX,2 )
q3
RX (ϕX,3 )
(1)
···
RX (ϕX,1 )
(D)
RY (ϕY,1 )
(1)
···
RX (ϕX,2 )
(D)
RY (ϕY,2 )
(1)
···
RX (ϕX,3 )
(D)
RY (ϕY,3 )
layer 1 upload
CZ ring
(D)
(D)
(D)
layer D upload
CZ ring
Fig. 3: Illustrative 3-qubit multi-depth QKS episode schematic showing only the first and final layer-specific RX -then-RY uploads. CZ-ring entanglers are inserted only between successive upload layers, so the displayed pattern is layer 1 upload → CZ ring → · · · → CZ ring → layer D upload. The special case D = 1 recovers the shallow template. a vector representation r(x) ∈ Rd constructed from it. In this study, we consider three such representation families: rraw (x) = vec(x),
(1)
rDCT (x) = vec C1:kf , 1:kt (x) ,
(2)
rPCA (x) = Pm vec(x).
(3)
Here, vec(·) flattens its matrix argument into a vector, C1:kf , 1:kt (x) denotes the upper-left kf × kt coefficient block of the 2-D DCT of x, and Pm denotes the projection onto the first m principal directions learned from the training set. The raw representation retains the full 400×400 spectrogram and flattens it into a vector in R160000 . The DCT family first applies a separable 2-D DCT (type II, with orthonormal normalization) [26] to the spectrogram, retains the upper-left kf × kt block of DCT coefficients, normalizes that block using training-derived statistics, and then flattens it. This retained block corresponds to low-index DCT coefficients, that is, coefficients capturing coarse, slowly varying structure over the 2-D spectrogram array along both the frequency-bin axis and the time-bin axis, rather than low RF frequencies in the original signal. In the benchmark, we use kf × kt ∈ {16 × 16, 64 × 64, 128 × 128, 256 × 256}. The Principal Component Analysis (PCA) family instead first flattens the spectrogram and then projects it onto an m-dimensional linear subspace, with m ∈ {32, 128, 256, 512, 2048}. These representation families compress the same signal in qualitatively different ways. DCT is a fixed transform that explicitly preserves coarse 2-D spectrogram structure, while PCA is a learned linear projection applied to flattened spectrograms and preserves high-variance directions without maintaining the original grid structure in the same explicit way. This distinction is central to the ablation: raw and DCT inputs become vectors only after flattening, whereas PCA already returns a vector representation. B. Multi-depth QKS featurizer Given a representation vector r(x) ∈ Rd , the QKS featurizer constructs one quantum feature block per episode. In line with the original QKS, multiple episodes are used to map the same
input to different randomized quantum states, and each such perepisode state is measured through a fixed observable family to produce a corresponding per-episode feature block. The shallow QKS template [19] corresponds to a single data-upload layer. Here, we extend that template to a depth-D feature map with repeated data re-uploading. For each episode e ∈ {1, . . . , E} and layer ℓ ∈ {1, . . . , D}, the model samples two affine maps from the representation space to the n input-dependent circuit angles, (e,ℓ)
(e,ℓ)
r(x) + bX ,
(e,ℓ)
(4)
(e,ℓ) (e,ℓ) (e,ℓ) ϕY (x) = WY r(x) + bY ,
(5)
ϕX (x) = WX (e,ℓ)
(e,ℓ)
(e,ℓ)
(e,ℓ)
with WX , WY ∈ Rn×d and bX , bY lently, for each qubit j ∈ {0, . . . , n − 1}, (e,ℓ)
(e,ℓ)
∈ Rn . Equiva-
ϕX,j (x) = ⟨wX,j , r(x)⟩ + bX,j ,
(e,ℓ)
(6)
(e,ℓ) (e,ℓ) (e,ℓ) ϕY,j (x) = ⟨wY,j , r(x)⟩ + bY,j .
(7)
(e,ℓ)
(e,ℓ)
In the completed benchmark, every entry of WX and WY is sampled i.i.d. from N (0, σ 2 ) with σ = 2, which was found (e,ℓ) empirically to yield good results, while every entry of bX (e,ℓ) and bY is sampled i.i.d. from U([0, 2π]). These angle vectors parameterize a layered circuit template. We define the single-layer data-upload operator as Ue,ℓ (x) =
n−1 Y
(e,ℓ) (e,ℓ) RY ϕY,j (x) RX ϕX,j (x) ,
(8)
j=0
and we denote by Vring either the identity (when entanglement is disabled) or the fixed controlled-Z (CZ) ring (0, 1), (1, 2), . . . , (n−2, n−1), (n−1, 0) otherwise. For episode e, the input state |0⟩⊗n is mapped to |ψe (x)⟩ = Ue (x)|0⟩⊗n ,
(9)
where Ue (x) = Ue,D (x) Vring · · · Ue,2 (x) Vring Ue,1 (x).
(10)
In other words, each episode re-uploads the same classical representation of the input across multiple randomized layers [27], [28], while entanglement, when enabled, is inserted only
between consecutive upload layers. This preserves a lightweight per-layer circuit structure while allowing the effective quantum feature map to become richer as depth increases. In the circuit schematic below, each upload block is drawn left-to-right as an RX gate followed by an RY gate, which corresponds to the operator product RY RX written above. The special case D = 1 recovers the standard shallow QKS template [19]. Figure 3 summarizes this multi-depth circuit structure. In our method, the measured observable family is taken to consist of all one- and two-body Pauli observables, n o On = σa(i) : 0 ≤ i < n, a ∈ {X, Y, Z} n o (j) ∪ σa(i) σb : 0 ≤ i < j < n, a, b ∈ {X, Y, Z} , (11) so that |On | = 9 n2 + 3n, where the factor 9 comes from the 3 × 3 possible choices of (a, b) ∈ {X, Y, Z} for each qubit pair (i, j). This observable family captures both single-qubit statistics and pairwise correlations induced by the feature map, while remaining polynomial in n. The per-episode feature block is therefore ξe (x) = ⟨ψe (x)|O|ψe (x)⟩ O∈O ∈ R|On | , (12)
setting. A linear SVM tests large-margin linear separability [29], logistic regression tests whether the same feature geometry supports a stable probabilistic linear decision rule [30], and the Random Fourier Features (RFF) and Nyström lifts test whether additional gains appear only after a controlled approximatekernel nonlinearity is applied [31], [32]. This distinction tells us whether the QKS features are already useful under simple linear models or whether their benefit emerges only after a richer nonlinear readout is added. The shortlisted representations are therefore evaluated with five matched readout families. For any Stage 5 readout ρ, let fρ : RE|On | → R denotes its score on ξ(x). We write ⊤ flinSVM (ξ) = wsvm ξ + bsvm ,
(14)
⊤ flog (ξ) = wlog ξ + blog , ⊤ fRff-SVM (ξ) = wrff-svm φRff (ξ) + brff-svm , ⊤ fRff-log (ξ) = wrff-log φRff (ξ) + brff-log , ⊤ fNys-log (ξ) = wnys φNys (ξ) + bnys .
(15) (16) (17) (18)
Here, φRFF and φNys denote fixed approximate-kernel feature lifts applied to ξ(x). For the logistic readouts, the corresponding positive-class probability is obtained by applying the sigmoid n and the final QKS feature vector is obtained by concatenating map to the score. For the approximate-kernel probes, we use matched feature lifts φRFF , φNys : RE|On | → Rm with all per-episode feature blocks as follows, projection dimension m = 1024, so these heads test whether ξ(x) = [ξ1 (x)⊤ , . . . , ξE (x)⊤ ]⊤ ∈ RE|On | . (13) QKS gains persist after a controlled nonlinear lift of fixed size The number of episodes E controls how many independently before the final linear score layer. In the final paired benchmark, sampled feature blocks contribute to ξ(x), while the depth D we compare each QKS model against a matched direct-readout controls how many data re-uploading and entanglement layers baseline, meaning that the same shortlisted representation vector r(x) is passed directly to the same readout family instead of are applied within each episode. first being mapped to ξ(x). The direct-readout baseline and the QKS model therefore share the same data partitions, the C. Five-stage protocol The benchmark is designed to disentangle the effects of same input representation, and the same readout family; only architecture, depth, depth–episode allocation, input represen- the feature map changes. b) Protocol: Stages 1–4 rank candidates by validation tation, and final readout, while keeping the raw test split AUROC first and validation F1 second. Stage 5 reports test untouched until the final evaluation. Stages 1–4 operate only AUROC and test F1 for the finalists, with F1 computed from on a fixed model-selection subset drawn from the raw training thresholded predictions at a fixed threshold of 0.5. split together with a disjoint validation subset. Stage 5 retrains the finalists on the full raw training split and evaluates them 1) Stage 1: shallow architecture search. Fix the reponce on the untouched raw test split. The completed benchmark resentation to the 64 × 64 DCT representation, set therefore contains 60 + 24 + 4 + 20 + 30 = 138 evaluations. depth to D = 1, and sweep n ∈ {2, 4, 6, 8, 10} qubits, For the benchmark reported here, the model-selection subset E ∈ {8, 16, 32, 64, 128, 256} episodes, and entanglement used in Stages 1–4 comprises approximately 20% of the full on/off. This stage identifies promising shallow operattraining split, obtained by stratified random sampling while ing points before any additional re-uploading depth is preserving coverage across abnormal groups; only Stage 5 uses introduced. the full training split. This reduced subset keeps the search 2) Stage 2: depth sweep. Take the top four Stage 1 tractable without collapsing anomaly diversity. configurations and sweep depth D ∈ {1, 2, 4, 6, 8, 10} a) Readouts: During Stages 1–4, we deliberately use a while keeping each candidate’s Stage 1 qubit count, single fast linear Support Vector Machine (SVM) readout so episode count, entanglement setting, and the 64×64 DCT that model selection remains focused on the quality of the representation fixed. This stage asks whether additional quantum feature map itself, rather than being confounded by re-uploading depth improves performance once a strong a broader comparison of downstream classifiers. In Stage 5, shallow configuration has already been identified. we then compare alternative corresponding readout families, 3) Stage 3: matched depth–episode trade-off. Retain as they probe different aspects of the same QKS feature only the single best Stage 2 family, meaning the geometry, and evaluate their performance under a matched same representation, qubit count, and entanglement
setting. Stage 3 does not keep the Stage 2 winner’s the randomized feature map does not necessarily translate into exact depth and episode count. Instead, it replaces better performance. Instead, moderate configurations appear that winning (D, E) pair with the matched alternatives sufficient to capture much of the discriminative structure, while (D, E) ∈ {(1, 256), (2, 128), (4, 64), (8, 32)}. This com- larger settings may introduce redundancy and additional modelparison therefore tests whether the observed Stage 2 gain selection variance. is attributable to depth itself or instead to a different b) Stage 2: depth sweep: Depth materially improves allocation between circuit depth and the number of performance once a strong shallow shell has been identified. episodes. Here, the four top performing models from Stage 1 are 4) Stage 4: representation comparison. Evaluate the evaluated, and the best Stage 2 model uses 10 qubits, 256 two strongest Stage 3 candidates across all ten input episodes, depth 4, and entanglement, reaching 0.8502 validation representations: raw input, four DCT variants, and five AUROC and 0.7718 validation F1 . On that same shell, moving PCA variants, after which the best QKS configuration is from depth 1 to depth 4 raises validation AUROC from 0.8282 retained for each shortlisted representation. This stage to 0.8502 and validation F1 from 0.7494 to 0.7718. Figure 4c compares representation families under QKS settings shows a non-monotonic depth trend across the four shortlisted already identified earlier, rather than reopening the full shells: each curve improves once some re-uploading is added, architecture search separately for every representation. but the strongest entangled q = 10, E = 256 configuration 5) Stage 5: final full dataset benchmark. For the fipeaks at depth 4 and then declines. The benchmark therefore nal shortlisted representations, compare matched directfavors a moderate re-uploading depth rather than the largest readout baselines and QKS models across the same five tested depth. This behavior is consistent with repeated data rereadout families on the full raw training split, and report uploading enriching the nonlinear feature map up to a moderate paired results on the held-out test data. depth, after which additional layers increasingly randomize or This design keeps the combinatorial search tractable while dilute task-relevant structure relative to the fixed linear readout. yielding interpretable answers to five distinct questions: which The fact that the strongest configurations are consistently shallow architectures are promising, whether depth helps on a entangled further suggests that interaction terms between qubits fixed shallow shortlist, whether the Stage 2 gain persists under provide useful additional structure beyond independent singlethe matched depth–episode trade-off, which input representa- qubit transformations. Overall, the results indicate that the best tions remain most effective under the shortlisted QKS settings, performance is obtained by combining entanglement with a and whether those gains persist under a final paired benchmark moderate re-uploading depth, while increasing circuit depth against matched direct-readout baselines. beyond this regime provides no systematic benefit and may degrade generalization. IV. R ESULTS c) Stage 3: matched depth–episode trade-off: Stage 3 A. Ablation study of pipeline design choices introduces a fixed-budget comparison in which the nominal Figure 4 summarizes the evidence from the model-selection product D × E = 256 is held constant while that budget is restages. Its four panels isolate the entangled Stage 1 sweep ( distributed between circuit depth and the number of episodes. It Fig. 4a), the unentangled Stage 1 sweep (Fig. 4b), the Stage 2 remains directly anchored to Stage 2 in that it keeps the winning depth-response curves (Fig. 4c), and the Stage 3 matched Stage 2 family fixed in terms of representation, qubit count, and depth–episode trade-off (Fig. 4d). entanglement, but it does not preserve the Stage 2 winner’s exa) Stage 1: shallow architecture search: On the fixed 64× act (D, E) = (4, 256) setting. Instead, it evaluates the matched 64 DCT representation, the best shallow model uses 10 qubits, alternatives (D, E) ∈ {(1, 256), (2, 128), (4, 64), (8, 32)} in 256 episodes, depth 1, and entanglement, reaching a validation order to test whether the Stage 2 gain is genuinely attributable AUROC of 0.828 and a validation F1 score of 0.749. The to depth or whether a similar budget is better spent on additional second largest alternative is a 8-qubit, 64-episode unentangled episodes. The strongest candidate is the 10-qubit, 64-episode, configuration, which reaches 0.826 AUROC and 0.729 F1 . depth-4 model, with 0.8482 validation AUROC and 0.7649 Because the Stage 1 ranking rule uses validation AUROC first validation F1 . The 10-qubit, 128-episode, depth-2 alternative and validation F1 second, the entangled 4-qubit, 64-episode remains competitive at 0.8397 AUROC and 0.7531, while model remains the Stage 1 winner despite that slightly higher the shallow (D, E) = (1, 256) and deep (D, E) = (8, 32) F1 . Figures 4a and 4b therefore suggest that the performance endpoints are weaker. Figure 4d makes the trade-off explicit: landscape is governed less by a broad monotonic dependence within this fixed-allocation comparison, (D, E) = (4, 64) leads on the number of qubits or QKS episodes than by a limited on both validation AUROC and validation F1 , (D, E) = set of favorable hyperparameter combinations. In the entangled (2, 128) remains the strongest alternative, and the shallow case, the best-performing configurations remain concentrated and very deep endpoints trail. The Stage 3 result suggests within a relatively compact region of the sweep, whereas the that the best performance comes from balancing per-episode unentangled case exhibits a flatter and less structured profile, expressivity and cross-episode diversity: too little depth leaves with improvements appearing more sporadically. This suggests each episode under-expressive, whereas too few episodes reduce that, for the fixed 64 × 64 DCT input, increasing the size of the averaging benefits of the randomized ensemble.
2
0.741
0.762
0.776
0.795
0.785
0.792
4
0.785
0.781
0.809
0.815
0.812
0.816
6
0.773
0.777
0.807
0.821
0.815
0.816
8
0.770
0.804
0.809
0.820
0.805
0.826
0.82
2
0.741
0.762
0.776
0.795
0.785
0.792
4
0.786
0.792
0.808
0.800
0.811
0.793
6
0.801
0.791
0.775
0.795
0.802
0.818
8
0.810
0.805
0.789
0.807
0.794
0.815
0.82 0.81
0.78 0.77
0.80
Validation AUROC
0.79
Qubits
0.80
Validation AUROC
Qubits
0.81
0.79 0.78 0.77
0.76 10
0.783
0.803
0.818
0.826
0.821
0.828
8
16
32
64
128
256
Episodes
0.76 10
0.75
(a) Entangled shallow QKS AUROC heatmap on fixed dct64x64.
0.816
0.794
0.778
0.793
0.788
0.818
8
16
32
64
128
256
Episodes
(b) Unentangled shallow QKS AUROC heatmap on fixed dct64x64.
0.85
Validation AUROC Validation F1
0.850
0.84
0.825
0.83
Validation score
Validation AUROC
0.75
0.82 0.81 0.80
q=10, E=256, ent=True q=8, E=256, ent=True q=10, E=64, ent=True q=6, E=64, ent=True
0.79 2
4
6
Depth
8
0.800 0.775 0.750 0.725 0.700
10
d=1 E=256
d=2 E=128
d=4 E=64
d=8 E=32
(c) Depth-response curves for the four shortlisted shallow configura- (d) Validation comparison across matched (D, E) pairs with fixed tions. depth–episode allocation.
Fig. 4: Validation-stage coreset results used to select finalists across Stages 1–3 of the ablation protocol.
d) Stage 4: representation comparison: When the two strongest Stage 3 QKS candidates are compared across all suggested input representations, the DCT family separates sharply from the remaining ones. The three strongest advancing representations are dct128x128, dct64x64,and dct256x256; their best QKS configurations all remain in the entangled 10-qubit regime. The best non-DCT contender, pca32, reaches only 0.5262 validation AUROC, well below the DCT advancing cutoff. Figure 5 makes this pattern explicit across all ten input representations, while Table II reports the three DCT finalists that advance to Stage 5, where the final QKS settings are fixed and for the final evaluation. Under this comparison, the benchmark does not support a representationagnostic QKS benefit; instead, it identifies a DCT-dominant regime in which structure-preserving classical preprocessing remains part of the effective hybrid model. This sharp DCT advantage is consistent with the idea that anomaly signatures
TABLE II: Stage 4 finalists that advance under the representation comparison and their locked QKS configurations. Repr.
q
Ep.
D
Ent.
Val AUROC
Val F1
dct128x128 dct64x64 dct256x256
10 10 10
64 64 128
4 4 2
Y Y Y
0.8541 0.8482 0.8152
0.7774 0.7649 0.7299
are organized in the time–frequency plane and are therefore better preserved by low-index DCT coefficients than by raw flattened inputs or PCA directions learned from global variance. B. Final benchmark on the full dataset Stage 5 retrains the Stage 4 finalists on the full raw training split and evaluates them once on the held-out raw test split. Among the 15 final representation–readout comparisons, the
0.85
Validation AUROC
0.80 0.75 0.70 0.65 0.60 0.55 0.50 0.45
8
dc
2 8x1 2 t1
4 4x6
6 dct
6
dc
5 6x2 5 t2
6
6x1
1 dct
32 pca
48
20 pca
Representation
6
25 pca
8
12 pca
2
51 pca
raw
0
x40
0 40
Fig. 5: Stage 4 comparison across input representations. Each bar reports the validation AUROC obtained by applying the two strongest Stage 3 QKS candidates across the ten representation families. DCT representations dominate, while PCA and raw inputs remain well below the advancing cutoff.
highest test AUROC is obtained by dct64x64 + LinearSvm, visibly smaller. This suggests that most of the benefit is already which reaches 0.8778 test AUROC and 0.7938 test F1 . The realized at the level of linear separability, leaving less residual highest test F1 is obtained by dct64x64 + LinearLogistic, structure for the approximate-kernel heads to recover. which reaches 0.7995 at a test AUROC of 0.8716. Figure 6 Third, this readout dependence is itself informative about summarizes the paired direct-readout-versus-QKS test AUROC the utility of the QKS features. The fact that the strongest comparison across all 15 representation–readout combinations, gains arise for the simplest linear readouts indicates that the while Table III reports the exact Stage 5 numbers for the six QKS mapping makes the shortlisted DCT representations more linear cases that exhibit the largest gains. The fact that the best linearly separable on the test split, rather than merely adding AUROC and best F1 arise from different DCT configurations redundancy that only richer nonlinear heads can exploit. In suggests that retaining more coefficients is most useful for practical terms, the QKS map appears to act primarily as ranking quality, whereas a more compact representation can a representation enhancer whose effect is strongest when the yield a cleaner decision boundary at the fixed threshold used downstream classifier is deliberately simple, while still yielding for F1 . a positive, but smaller, lift for the nonlinear readouts. Three key patterns can be identified from these results. First, C. Validation on real quantum hardware all 15 rows in Fig. 6 place the QKS pipeline above its matched To assess the hardware-realism aspect of the study, we direct-readout baseline. Here, the direct-readout baseline refers to the same input representation and the same readout family performed a focused validation on quantum hardware. Due applied directly to the representation vector r(x), without the to the cost of real-device execution, all experiments were QKS feature map. The observed AUROC gain is therefore conducted with an entangled feature map using 4 qubits, a systematic within this finalized benchmark rather than confined single-layer depth, and 32 episodes, with expectation values to a single isolated representation–readout pair, indicating that estimated from 4096 shots per circuit. These experiments were th th the QKS map consistently reshapes the shortlisted DCT inputs run on the ibm_quebec QPU from 25 to 29 March 2026, into a more discriminative geometry than the matched direct- using the dct128x128 representation. In contrast to the completed simulation study, the measured observable family readout pipeline. was restricted to all single-qubit Z observables and pairwise Second, the reporting ∆AUROC and ∆F1 , which quantify ZZ observables, the difference between QKS and direct-readout baseline scores, (n) (i) (i) (j) receptively, show that the largest gains arise for the linear OZ,ZZ = {σZ }n−1 (19) i=0 ∪ {σZ σZ }0≤i<j<n , readouts. Across the six linear rows in Table III, the AUROC (n) gains range from +0.0952 to +0.1839, with the largest gap so that |OZ,ZZ | = n(n+1) , instead of the full one- and 2 attained by dct64x64 + LinearSvm. The approximate- two-body Pauli family considered in simulation. In addition, kernel readouts also improve, but their AUROC gains are the dataset size was reduced to approximately 10% of the
dct64x64 | LinearSvm dct128x128 | LinearSvm dct256x256 | LinearSvm dct64x64 | LinearLogistic dct128x128 | LinearLogistic dct256x256 | LinearLogistic dct64x64 | NystroemLogistic dct64x64 | RffLogistic dct128x128 | NystroemLogistic dct256x256 | NystroemLogistic dct64x64 | RffLinearSvm dct128x128 | RffLogistic dct128x128 | RffLinearSvm dct256x256 | RffLogistic dct256x256 | RffLinearSvm
Direct-readout test AUROC QKS test AUROC
0.675
0.700
0.725
0.750
0.775
0.800
Test AUROC
0.825
0.850
0.875
Fig. 6: Stage 5 paired test AUROC comparison between matched direct-readout baselines and QKS pipelines. Each row corresponds to one representation–readout comparison on the untouched raw test split after the locked finalists from a single finalized benchmark run are retrained on the full raw training split. The largest AUROC gaps appear for the linear readouts, while the approximate-kernel readouts also benefit from QKS but by smaller margins. TABLE III: Compact Stage 5 benchmark slice for the linear readouts. Repr.
Readout
dct64x64 dct64x64 dct128x128 dct128x128 dct256x256 dct256x256
LinearSvm LinearLogistic LinearSvm LinearLogistic LinearSvm LinearLogistic
Direct AUROC
QKS AUROC
∆AUROC
Direct F1
QKS F1
∆F1
0.6939 0.7266 0.6945 0.7294 0.6975 0.7312
0.8778 0.8717 0.8763 0.8705 0.8576 0.8264
0.1839 0.1451 0.1818 0.1411 0.1601 0.0952
0.6515 0.6607 0.6519 0.6614 0.6558 0.6635
0.7938 0.7995 0.7969 0.7978 0.7754 0.7739
0.1423 0.1389 0.1451 0.1363 0.1195 0.1104
original training and test sets. The search space was deliberately restricted so as to remain within the available quantum budget and to limit the practical overhead associated with queue times and repeated hardware calls. As a result, this experiment was not intended to reproduce the full simulation study, but rather to evaluate the practical viability of the approach under genuine real-device conditions. It should therefore be viewed as a focused real-QPU benchmark aimed at assessing whether the learned representation retains meaningful predictive structure when executed on an actual superconducting quantum processor. The resulting comparison between real-QPU and simulator performance is reported in Table IV for several classical readout models, with the simulator baseline obtained by re-running the pipeline under exactly the same configuration used on real-QPU—identical feature map, number of qubits, circuit
(n)
depth, number of episodes, observable family OZ,ZZ , and train/test splits—so that the execution backend is the only varying factor. Overall, the real-hardware results remain close to their simulated counterparts, with only small deviations in AUROC across all tested readouts. Here, HS denotes the hardware–simulator difference: ∆HS AUROC = AUROCQPU − AUROCsim . More specifically, the observed AUROC gaps range from −1.281 × 10−2 to +7.132 × 10−3 , indicating that the ranking quality of the QKS features is largely preserved under realdevice execution despite the reduced measurement setting. The F1-score differences, ∆HS F1 = F1QPU − F1sim ,
Readout LinearSvm LinearLogistic RffLinearSvm RffLogistic NystroemLogistic
Hardware AUROC
Hardware F1
Simulator AUROC
Simulator F1
∆HS AUROC
∆HS F1
0.738302 0.720810 0.807438 0.803232 0.813637
0.605144 0.658851 0.719892 0.736967 0.748166
0.731170 0.733619 0.805798 0.805961 0.815457
0.657754 0.645161 0.704762 0.702247 0.716418
+7.132 × 10−3 −1.281 × 10−2 +1.640 × 10−3 −2.729 × 10−3 −1.819 × 10−3
−5.261 × 10−2 +1.369 × 10−2 +1.513 × 10−2 +3.472 × 10−2 +3.175 × 10−2
TABLE IV: Comparison of real-QPU and simulator performance across classical readout models, both evaluated under identical settings: 4-qubit entangled feature map, single-layer depth, 32 episodes, OZ,ZZ observables, train/test split. are somewhat more variable, but remain favorable for most readouts. In particular, the NystroemLogistic readout achieves the best real-QPU performance, with an AUROC of 0.813637 and an F1 score of 0.748166, compared with 0.815457 and 0.716418 in simulation. Taken together, these results indicate that, within this deliberately reduced experimental setting, the proposed pipeline remains robust to real-device execution and preserves a substantial fraction of the predictive structure observed in simulation. V. D ISCUSSION AND L IMITATIONS a) Multi-depth QKS acts as a representation enhancer: The completed benchmark does not suggest that QKS uniformly dominates classical baselines. Instead, it shows that repeated data re-uploading can improve separability in specific representation regimes, with the strongest gains appearing once depth is increased beyond the shallow template but kept below the largest tested values. Entanglement also appears to contribute positively in the strongest regimes: entangled configurations dominate among the top-performing shallow models and outperform their unentangled counterparts in most matched comparisons, although this benefit is not systematic across the entire search space. Taken together, these results suggest that the strongest QKS representations arise from a combination of a suitable input representation, moderate reuploading depth, and useful inter-qubit interactions, rather than from indiscriminately increasing circuit complexity. b) The strongest regime is DCT-specific: The featurerepresentation stage is decisive: all top-performing configurations use DCT variants, whereas PCA and raw inputs do not remain competitive. This suggests that, for spectrogrambased RF anomaly detection, performance depends not on the QKS feature map alone, but on its combination with a suitable classical representation. Our interpretation is that DCT is more effective because it preserves the structured organization of anomaly patterns in the time–frequency plane while concentrating informative content into a compact set of coefficients. By contrast, PCA emphasizes directions of maximal global variance, which can mix distant time–frequency regions and obscure features that are more relevant for anomaly discrimination. Raw inputs, in turn, likely preserve excessive irrelevant variability and redundancy, thereby reducing the ability of the QKS feature map to produce a representation aligned with the discriminative structure of the task. c) Readout dependence remains informative: All final readouts benefit from QKS, yet the largest paired gains appear
for the linear probes. This is informative as it shows that QKS is changing the geometry seen by simple classical classifiers, not merely adding redundancy that only richer downstream models can exploit. In practical terms, this is the main value of the QKS features in this study: they act as a representation enhancer whose effect is largest when the downstream classifier is deliberately simple, while still yielding a positive, but smaller, lift for the nonlinear readouts. d) Real-device execution preserves feature quality: The hardware study confirms that QKS features retain meaningful predictive structure under real-device conditions. Although the QPU experiment uses a reduced configuration with fewer qubits, shallower depth, a restricted observable set, and a smaller dataset, the results remain within 0.013 AUROC of their simulated counterparts across all tested readouts. Notably, the ranking order across readouts is preserved between hardware and simulation, with NystroemLogistic achieving the best performance in both settings. This consistency indicates that the feature geometry learned by the QKS pipeline is not critically disrupted by hardware noise at this scale, and supports the practical viability of the approach on near-term superconducting processors. We emphasize that this experiment was not designed to reproduce the full simulation benchmark, but to assess whether the pipeline remains functional under genuine realdevice constraints, a question it answers affirmatively. Taken together, these observations suggest that the study probes realism along both the data and hardware axes, placing the benchmark closer to realistic hybrid deployment conditions. e) Limitations: The evidence remains task-specific. The dataset is controlled rather than drawn from a live operational RF deployment, and generalization to unseen interference types or real-world spectrum conditions has not been tested. The completed study uses one finalized run rather than a broad seed ensemble, so stability across random initializations of the affine maps has not been quantified. The classical readout family is intentionally restricted to fast probes and kernel approximations, and stronger classical heads may reduce or eliminate the observed QKS gains. We therefore make no universal claim about QKS on every anomaly detection task, and we do not interpret the present results as evidence of quantum advantage. VI. C ONCLUSION This paper studied Quantum Kitchen Sinks for RF spectrogram anomaly detection under a controlled, leakage-free evaluation protocol. Starting from the standard shallow QKS
template, we introduced a multi-depth variant with repeated data re-uploading and optional ring entanglement, and evaluated it on spectrograms built from real measured sub-6 GHz cellular signals with synthetically injected anomalous transmissions. The resulting benchmark was designed not only to compare models, but also to identify which components of the hybrid pipeline matter most for this task. The results support a clear picture within this benchmark. QKS does not provide a representation-agnostic improvement over classical baselines; instead, its utility depends strongly on the interaction between the quantum feature map, the input representation, and the downstream readout. The strongest results arise in DCT-based regimes, where moderate-depth entangled QKS configurations outperform their matched directreadout baselines across all evaluated final readout families on held-out test data, reaching a best test AUROC of 0.8773 and a best test F1 of 0.8032. The same favorable pattern is not reproduced with raw or PCA-based inputs, and the largest gains do not come from the largest tested models. In addition, the reduced real-device validation on the ibm_quebec QPU preserves the main predictive trends seen in simulation, with AUROC deviations below 0.013. Future work should extend the study to additional RF anomaly families and test whether the same operating regimes persist under broader hardware-aware execution constraints. VII. ACKNOWLEDGEMENT The authors gratefully acknowledge the financial support provided by Defence Research and Development Canada (DRDC), which made this research possible. R EFERENCES [1] X. Jiang, P. Li, B. Li, Y. Zou, and R. Wang, “Intelligent Jamming Strategies for Secure Spectrum Sharing Systems,” IEEE Trans. Commun., vol. 70, no. 2, pp. 1153–1167, 2022. [2] Y. Zou, J. Zhu, X. Wang, and L. Hanzo, “A Survey on Wireless Security: Technical Challenges, Recent Advances, and Future Trends,” Proc. IEEE, vol. 104, no. 9, pp. 1727–1765, 2016. [3] A. S. Ali, W. T. Lunardi, G. Singh, L. Bariah, M. Baddeley, M. A. Lopez, J.-P. Giacalone, and S. Muhaidat, “RF Jamming Dataset: A Wireless Spectral Scan Approach for Malicious Interference Detection,” IEEE Commun. Mag., vol. 62, no. 11, pp. 114–120, 2024. [4] M. Lichtman, R. P. Jover, M. Labib, R. Rao, V. Marojevic, and J. H. Reed, “LTE/LTE-A Jamming, Spoofing, and Sniffing: Threat Assessment and Mitigation,” IEEE Commun. Mag., vol. 54, no. 4, pp. 54–61, 2016. [5] M. Lichtman, J. D. Poston, S. Amuru, C. Shahriar, T. C. Clancy, R. M. Buehrer, and J. H. Reed, “A Communications Jamming Taxonomy,” IEEE Secur. Priv., vol. 14, no. 1, pp. 47–54, 2016. [6] H. Pirayesh and H. Zeng, “Jamming Attacks and Anti-Jamming Strategies in Wireless Networks: A Comprehensive Survey,” IEEE Commun. Surv. Tutor., vol. 24, no. 2, pp. 767–809, 2022. [7] X. Wang, J. Wang, Y. Xu, J. Chen, L. Jia, X. Liu, and Y. Yang, “Dynamic Spectrum Anti-Jamming Communications: Challenges and Opportunities,” IEEE Commun. Mag., vol. 58, no. 2, pp. 79–85, 2020. [8] S. Rajasegarar, C. Leckie, and M. Palaniswami, “Anomaly Detection in Wireless Sensor Networks,” IEEE Wireless Commun., vol. 15, no. 4, pp. 34–40, 2008. [9] T. Yucek and H. Arslan, “A Survey of Spectrum Sensing Algorithms for Cognitive Radio Applications,” IEEE Commun. Surv. Tutor., vol. 11, no. 1, pp. 116–130, 2009. [10] A. Krause, M. D. Khursheed, P. Schulz, F. Burmeister, and G. Fettweis, “Digital Twin of the Radio Environment: A Novel Approach for Anomaly Detection in Wireless Networks,” in 2023 IEEE Globecom Workshops (GC Wkshps). IEEE, 2023, pp. 1307–1312.
[11] Z. Wu, Y. Zhao, Z. Yin, and H. Luo, “Jamming Signals Classification Using Convolutional Neural Network,” in 2017 IEEE International Symposium on Signal Processing and Information Technology (ISSPIT). IEEE, 2017, pp. 062–067. [12] H. Xu, X. Ma, C. Wang, X. Wang, C. Xu, F. Gao, and L. Kong, “A neural network approach for wireless spectrum anomaly detection in 5G-unlicensed network,” CCF Transactions on Pervasive Computing and Interaction, vol. 4, no. 4, pp. 465–473, 2022. [13] S. Rajendran, W. Meert, V. Lenders, and S. Pollin, “Unsupervised Wireless Spectrum Anomaly Detection With Interpretable Features,” IEEE Trans. Cogn. Commun. Netw., vol. 5, no. 3, pp. 637–647, 2019. [14] N. Tandiya, A. Jauhar, V. Marojevic, and J. H. Reed, “Deep Predictive Coding Neural Network for RF Anomaly Detection in Wireless Networks,” in 2018 IEEE International Conference on Communications Workshops (ICC Workshops). IEEE, 2018, pp. 1–6. [15] J. Ren, X. Zhang, and Y. Xin, “Using Deep Convolutional Neural Network to Recognize LTE Uplink Interference,” in 2019 IEEE Wireless Communications and Networking Conference (WCNC). IEEE, 2019, pp. 1–6. [16] X. Zhou, J. Xiong, X. Zhang, X. Liu, and J. Wei, “A Radio Anomaly Detection Algorithm Based on Modified Generative Adversarial Network,” IEEE Wireless Commun. Lett., vol. 10, no. 7, pp. 1552–1556, 2021. [17] P. Liao, X. Wang, Y. Shan, L. An, and S. Mao, “Wireless Sensing in Artificial Intelligence of Things: A General Quantum Machine Learning Framework,” IEEE Netw., vol. 39, no. 3, pp. 207–214, 2025. [18] I. Ahmad, L. Yang, Y. Alkhrijah, A. Almadhor, M. A. Alawad, L. Peng, and P.-H. Ho, “Quantum Machine Learning for Anomaly Detection: The Future of Smarter and Safer IoT Networks,” IEEE Netw., vol. 40, no. 2, pp. 30–38, 2026. [19] C. M. Wilson, J. S. Otterbach, N. Tezak, R. S. Smith, G. E. Crooks, and M. P. da Silva, “Quantum kitchen sinks: An algorithm for machine learning on near-term quantum computers,” 2018. [Online]. Available: https://arxiv.org/abs/1806.08321 [20] M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincio, and P. J. Coles, “Variational quantum algorithms,” Nature Reviews Physics, vol. 3, no. 9, pp. 625–644, 2021. [21] E. Mohammadisavadkoohi, N. Shafiabady, and J. Vakilian, “A Systematic Review on Quantum Machine Learning Applications in Classification,” IEEE Trans. Artif. Intell., vol. 7, no. 1, pp. 4–21, 2026. [22] J. Kim, H. Kim, and B. Kim, “Wireless Anomaly Signal Dataset (WASD): An Open Dataset for Wireless Cellular Spectrum Monitoring and Anomaly Detection,” IEEE Access, vol. 12, pp. 196 240–196 248, 2024. [23] J. Kim, S. Oh, H. Kim, and B. Kim, “Spectrum Anomaly Detection Using Deep Neural Networks: A Wireless Signal Perspective,” IEEE Access, vol. 13, pp. 154 948–154 959, 2025. [24] P. Lohan, B. Kantarci, M. A. Ferrag, N. Tihanyi, and Y. Shi, “From 5G to 6G Networks: A Survey on AI-Based Jamming and Interference Detection and Mitigation,” IEEE Open J. Commun. Soc., vol. 5, pp. 3920–3974, 2024. [25] B. Kim, S.-H. Kong, and S. Kim, “Low Computational Enhancement of STFT-Based Parameter Estimation,” IEEE J. Sel. Top. Signal Process., vol. 9, no. 8, pp. 1610–1619, 2015. [26] G. Wallace, “The JPEG Still Picture Compression Standard,” IEEE Trans. Consum. Electron., vol. 38, no. 1, pp. xviii–xxxiv, 1992. [27] A. Pérez-Salinas, A. Cervera-Lierta, E. Gil-Fuster, and J. I. Latorre, “Data re-uploading for a universal quantum classifier,” Quantum, vol. 4, p. 226, 2020. [28] M. Schuld, R. Sweke, and J. J. Meyer, “Effect of data encoding on the expressive power of variational quantum-machine-learning models,” Physical Review A, vol. 103, p. 032430, 2021. [29] C. Cortes and V. Vapnik, “Support-Vector Networks,” Machine Learning, vol. 20, no. 3, pp. 273–297, 1995. [30] D. R. Cox, “The Regression Analysis of Binary Sequences,” Journal of the Royal Statistical Society: Series B (Methodological), vol. 20, no. 2, pp. 215–232, 1958. [31] A. Rahimi and B. Recht, “Random Features for Large-Scale Kernel Machines,” in Advances in Neural Information Processing Systems 20. Curran Associates, Inc., 2007, pp. 1177–1184. [32] C. K. I. Williams and M. Seeger, “Using the Nyström Method to Speed Up Kernel Machines,” in Advances in Neural Information Processing Systems 13. MIT Press, 2001, pp. 682–688.