Screaming Channel: Recent Advances in Far Field EM Side-Channel Attacks based on Radio Coupling

Abstract

The Internet of Things (IoT) is revolutionizing society by enabling anytime, anywhere connectivity, but its widespread deployment also makes IoT devices attractive targets for malicious activities. Over the past decade, radio-coupling based Far-field EM Side-Channel Attacks (FEM-SCAs), also known as screaming channel attacks, have emerged as a new threat to these wireless IoT edges. Unlike conventional power or near-field EM side-channel attacks, FEM-SCAs exploit the unintended coupling between digital computation and the Radio Frequency (RF) transmission chain, allowing sensitive information to be captured through radiated radio signals at long distances. Since the first screaming-channel demonstration, this field has progressed rapidly: attacks have become more efficient, Deep-Learning (DL) enabled methods have been developed, and recent studies have begun to examine more realistic wireless-protocol settings. Despite existing practical cases of remote FEM-SCA attacks, the literature remains fragmented. To systematically understand this security threat and to aid future countermeasure design, this paper presents the first systematic survey of existing screaming channel attack methodologies and case studies. We organize studies into a hierarchical framework covering leakage discovery and understanding, enhanced attack capability, attack targets and variants, and public data collections. Through a fine-grained cross-paper analysis, we compare representative works in terms of target implementation, attack distance, analysis method, profiling requirement, acquisition environment, and other factors. Building on this end-to-end review, we provide a focused discussion that distills FEM-SCA research into one central trade-off, three major limitations, and three future research roadmaps. Taken together, this survey offers a structured framework for understanding a rapidly developing far-field side-channel attack class, and provides practical guidance for evaluating its real-world threat boundaries [23], [24], [38], [85], [92], [99].

I. Introduction: From Digital Leakage to Radio

Side-Channel Attacks (SCAs) have become an increasingly prominent threat to the security of modern computing systems, particularly with the rapid expansion of embedded edge device deployment. The very silicon hearts of electronic devices, from sensors to controllers, may disclose sensitive information through physical leakage such as power fluctuations, electromagnetic (EM) emissions, timing variations, and others. By exploiting this unintentional physical leakage, adversaries can eventually recover secret data. Over the past two decades many side channels have been investigated, including compromising cryptographic implementations [1]–[3], reverse-engineering neural networks, stealing intellectual property, monitoring browser information, predicting random number generators, tracking code execution, private user identities, and file systems. Traditional SCAs are typically divided according to the leakage source they exploit: power analysis observes data-dependent power consumption [18], timing analysis measures execution-time variations [19], near-field EM analysis captures localized radiation [20], and optical and acoustic analyses exploit further channels [21], [22]. Although these side channels have been highly effective in laboratory and evaluation settings, their threat model is usually constrained by the measurement interface: power attacks often require physical contact with the chip, near-field EM attacks require probes placed close to the board, and some require visibility of the die. Conventional SCAs are therefore generally constrained by the need for physical proximity, which limits their practicality in scenarios where the attacker cannot obtain physical access to the victim device.

In 2018, a new type of side channel was discovered, called Far-field EM Side-Channel Attacks (FEM-SCAs), also known as Screaming Channel attacks, introduced by Camurati et al. at the 2018 CCS conference [23]. This is considered a growing threat as the side-channel attack can be conducted remotely, without physical proximity to the target device. In the novel FEM-SCAs, by exploring the coupling effect between EM emissions from the CPU core and radio carrier signals on mixed-signal chips, adversaries can eliminate the requirement of access to the victim device, which makes the attack increasingly covert (as shown in Fig. 1). Considering that the core components of widely deployed IoT edge devices are predominantly mixed-signal chips, the security threats posed by FEM-SCAs might be broader than expected. This attack method capitalizes on radio transmitters integrated into chips, permitting sensitive data to inadvertently modulate radio frequency signals, which are subsequently disseminated through the airwaves. Practical attack cases have been demonstrated remotely to successfully compromise Bluetooth device implementations of AES [23], [24], eavesdrop on clear-text data [25], and recover the key from protected crypto [26].

While some reviews on side-channel attacks exist [27]–[32], their coverage remains incomplete for screaming channels: they covered conventional approaches such as Differential Power Analysis (DPA) [18], Template Attacks (TAs) [34], Correlation Power Analysis (CPA) [35], and Test Vector Leakage Assessment (TVLA) [36], then Machine-Learning (ML) attacks [27], Deep-Learning SCAs (DLSCAs) [29], timing analysis [30], PQC attacks [31], and the Branch Prediction Unit (BPU) [32]. In 2026, [37] provided a high-level enumeration of screaming-channel-related studies without technical detail or systematic taxonomy. This paper therefore adopts a selective approach confined to works published in high-reputation conferences and journals, and narrows the scope to attack cases based on RF-coupled screaming channels with a capturing frequency located at Ultra High Frequency (UHF), excluding other remote side-channel attacks such as TEMPEST [42], [43], EMIRIS [44], and remote EM attacks [45] whose emanations are not coupled to an RF communication carrier. The contributions are: the first comprehensive survey dedicated to FEM-SCAs covering four dimensions (discovery and understanding, enhanced capability and practicality, targets and variants, and public data collections); a fine-grained cross-work analysis in terms of target implementation, attack distance, attack method, profiling requirement, key-enumeration complexity, and acquisition environment; and a boundary-aware discussion identifying one key trade-off, limitations, and future directions. The remainder is organized as follows. Section II introduces background on side channels and deep learning, Section III explains the theoretical foundations of radio coupling and remote capture, Section IV reviews early discovery and understanding, Section V summarizes enhanced attack capability, Section VI discusses targets and variants, Section VII presents public data collections, Section VIII gives open challenges and research directions, and Section IX concludes the survey.

II. Background

Side-channel attacks aim to recover sensitive information from electronic devices by exploiting the unintentional and unavoidable physical leakage, such as power consumption and EM emissions, generated during the execution of the target device [46], [47]. They are commonly classified into non-profiled and profiled: the former directly compromise cryptographic implementations with techniques like DPA [18] or CPA [35], while the latter first learn a leakage profile connecting captured side-channel measurements to the key-dependent sensitive intermediate value of the algorithm, then use it for the attack. In most cases profiled attacks demonstrate superior performance under the same conditions; for example, a deep-learning based power analysis (in the profiled setting) can recover the subkey of an AES implementation on an ATXmega128D4 MCU with only a single trace, while CPA typically demands approximately 50 traces for a comparable success rate [48]. Setting up profiled attacks typically involves significantly more preparation costs, however, since the adversary must fully control copies of the victim device (profiling devices) and capture extensive traces to build the leakage profile, as in template attacks [34], which use profiling-device traces to generate probability distributions that characterize traces by the key-dependent intermediate value.

In the last two decades, advances in deep learning have made SCAs significantly more threatening, predominantly employed in profiled settings [10], [51]–[55], where a well-trained deep-learning model can make the attack several orders of magnitude more efficient than conventional signal processing [56]–[59]. Various neural networks are increasingly used by adversaries to analyze traces, including Multi-Layer Perceptrons (MLPs), Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM), Transformers, and even Large Language Models (LLMs) [60]–[63]. For example, CNNs were first utilized to overcome jitter-based countermeasures [64], then reinforcement learning for hyperparameter tuning [65], then neural networks to augment template attacks [66]. Deep neural networks have even been successfully employed to compromise FPGA implementations of masked Kyber-512 as demonstrated in 2023 [67], Zaid et al. proposed a Conditional Variational Autoencoder in TCHES bridging DL models and SCA paradigms [68], a Broad Learning System (BLS) [69], [70] was used to speed up training, and a Multi-point Cluster-based (MC) labeling method succeeded without knowledge of plaintext or ciphertext [71].

III. Theoretical Foundations of Screaming Channels

Prior to the discovery of screaming channels, covert data recovery via a radio transmitter was implemented using amplitude-modulating analog/RF hardware Trojans in IEEE 802.11a/g WiFi transceivers [72], stealthy modifications creating a covert channel while remaining compliant with wireless standards [73]; FEM-SCAs share the same threat model of unauthorized radio transmission but constitute a fundamentally different attack vector: screaming channels are a side-channel attack manipulating a device's intentional transmitter by exploiting unintentional EM emissions from its digital components, requiring no hardware modifications. Unlike conventional side-channel attacks, FEM-SCAs eliminate the proximity constraint, enabling remote compromise of cryptographic implementations. This becomes critical when the victim embeds mixed-signal circuits integrating a noise-sensitive analog part (e.g., Bluetooth modules) with a digital part (e.g., CPU core) on a single die [74]. Digital circuits like microcontrollers or cryptographic accelerators rapidly switch transistors, generating sharp voltage and current variations correlated with the processed data, especially at the system clock frequency and its harmonics. Although ideally confined to the digital domain, this noise leaks through the silicon substrate via resistive or capacitive coupling [75] due to the tight physical proximity of analog components, particularly radio transceivers, in a phenomenon called spectrum spraying: the CPU core generates square-wave noise whose harmonics spread across the spectrum, and when these harmonics interact with data-dependent leakage signals they effectively modulate the harmonics with sensitive internal states. The modulated signals then couple into the analog part, where the RF block and Voltage-Controlled Oscillator (VCO) further modulate them to a radio frequency defined by the wireless protocol and transmit them via the on-chip antenna, to be captured remotely by adversaries to derive secrets — this is the FEM-SCA, or screaming channel attack. What makes it especially threatening is that leakage is not confined to near-field EM emissions [76] but the mixed-signal path acts as amplifier and transmitter, broadcasting leaked modulated signals as part of regular RF output and allowing extraction of AES keys from far-field EM at a distance using only passive listening equipment such as a radio. Since such leakage is a natural byproduct of placing digital and analog components on the same die, the screaming channel represents not a vendor-specific flaw but a design-level vulnerability with broad implications for RF-enabled device security.

Mathematically, the paper describes generation in time and frequency domains: the clock signal s1(t) from the computational core with square-wave frequency fs (Formula 1), its Fourier transform S1(f) with coefficients An and impulse function (Formula 2), An defined by duty cycle τ (Formula 3). The secret-dependent leakage s2(t) from the crypto block is then modulated by the clock square wave to yield s3(t) (Formula 4) and its spectrum S3(f) via convolution with S1 (Formula 5). Afterwards the RF block modulates the signal to high frequency with carrier ei2πfct at carrier frequency fc (Formula 6: s4(t) and S4(f)), then amplifies and transmits it through the on-chip antenna. At the receiver, Formula 7 shows the captured signal after demodulation with center receiving frequency fc + Nfs and front-end gain G: r(t) consists of a scaled version of the leakage s2(t) at baseband plus modulated copies at frequencies (n−N)fs, and a low-pass filter can extract s2(t) from the received signal. A typical screaming-channel scenario assumes the adversary can passively capture the signal at a specific radio spectrum and knows or infers the target's CPU clock frequency, since leakage appears at a predictable offset of the radio carrier frequency, allowing the receiver to be tuned to a harmonic where the side-channel signal is strong. The paper uses the Measurements to Disclosure (MTD) metric, the number of traces an attack requires to recover the secret key (i.e., PGE = 0).

IV. Discovery and Understanding

In 2018, Camurati et al. [23] conducted the first FEM-SCA case at the CCS'18 conference as the core discovery phase of the screaming channel: they revealed that cryptographic computations executed on mixed-signal SoCs can unintentionally leak information through the RF transmission path when digital switching activity couples into the radio front-end, fundamentally challenging the conventional assumption that EM side-channel attacks require near-field access [77] and instead demonstrating that sensitive information can be extracted from legitimate wireless emissions at distances. Importantly, this leakage does not originate from a dedicated side-channel radiation source but from unintended modulation of an already existing communication carrier, thereby enabling a new class of side channel. Using a Software-Defined Radio (SDR) [78], they captured EM leakage of TinyAES implementations [79] over distances of up to 10 m, strong enough for correlation analysis and template attacks to recover AES keys. They conducted template attacks in two capturing environments: a normal office at 1 m distance, and an anechoic chamber allowing measurements up to 10 m. Table I presents the full key recovery results: in the office at 1 m the template attack requires around 52,589 traces (×500 encryption repetitions) with key enumeration 20 and 70K×500 profiling traces, indicating effectiveness but lower signal-to-noise ratio; in the controlled laboratory only 718 traces suffice at 3 m, remaining highly feasible at 10 m with around 1,428 traces (and 130K×500 templates). Although [23] brought screaming channels to the forefront, many aspects remained insufficiently understood: deeper characterization of leakage sources, attack strategies fully exploiting this unconventional channel, and comprehensive security evaluations under realistic deployment scenarios.

Building on this, the follow-up TCHES'20 work [24] thoroughly characterized this novel attack vector: how leakage coexists with intended radio transmissions, how distance and frequency affect leakage strength and distortion, and whether profiling models can be reused across devices and environments. They tested the same template-attack method as in [23] for both controlled chambers and realistic office environments, added conventional EM capture support for comparison, and experimented on a real authentication protocol (Google Eddystone), employing multiple Nordic nRF52832 instances and different SDRs (HackRF, Ettus N210 USRP, B210 USRP) with antennas and amplifiers. Results show screaming-channel leaks are weaker and less structured than conventional ones but still viable under realistic conditions, that the choice of harmonic channel frequency does not significantly alter leakage strength or shape, and that far-field signal strength decreases quadratically with distance, requiring stronger amplification and better equipment. Experimentally, [24] showed that a correlation profile created with a cabled setup can be reused months later to recover TinyAES keys from another device at 15 m in an office (5K×1K traces with key enumeration up to 2^23), and that leakage remains detectable even at 60 m (Table II). A subsequent KTH thesis [81] complemented this understanding by systematically comparing power and far-field EM settings, showing learned leakage patterns are highly sensitive to device diversity, distance, and acquisition conditions. The CARDIS'24 work [82] further refined spectrum-level understanding by showing exploitable information also emerges in non-harmonic frequency regions, extending the frequency-domain interpretation of screaming channels to a broadband, modulation-dependent spectral phenomenon. In 2024, BlueScream [80] demonstrated that leakage is not limited to synthetic firmware but can be triggered and exploited in real-world Bluetooth Low Energy (BLE) stacks through careful manipulation of transmission timing and cryptographic execution windows: by manipulating low-level BLE protocol parameters, adversaries force a victim device to perform the critical Session Key (SK) derivation concurrently with a radio transmission, creating a far-field side channel allowing recovery of the Long Term Key (LTK) from up to 1 m away with a profiled correlation attack (Table III: 10K×300 down to 16K×1 traces), compromising the entire trusted pairing and decrypting all past and future communications. The review [37] adds an applicability-limits perspective at RFSoC level: FEM-SCA applicability is determined not only by Bluetooth usage or software AES but by whether the internal chip architecture provides a strong enough coupling path between sensitive computation and the RF module, with practical limits including difficulty identifying dominant leakage sources inside RFSoCs and the lack of screaming-channel resistance evaluation indicators in cryptographic security standards. Together these works establish a four-stage evolution (Fig. 5): core discovery, mechanism understanding, spectrum understanding (carrier, sidebands, harmonic and non-harmonic components), and boundary understanding where dense legitimate signals increasingly mask leakage.

V. Enhanced Attack Capability and Practicality

The evolution of screaming channel attacks has not only expanded theoretical understanding but also significantly enhanced practical attack capability in realistic cases; this section focuses on FEM-SCA enhancement techniques improving efficiency, robustness, and generalization. For fair comparison most strategies are evaluated on the same target: the TinyAES implementation. In 2020, deep-learning techniques were first introduced into screaming channel attacks [38], exploring to what extent a CNN model can make far-field EM side-channel attacks more efficient, and proposing a multi-distance, multi-device generalization strategy: on a Nordic Semiconductor nRF52832 SoC target supporting Bluetooth 5 [84] at 2 Mbps data rate, with an Ettus Research USRP N210 receiver and a Grid Parabolic Antenna TL-ANT2424B attached at center receiving frequency 2.528 GHz (i.e., 2fclock + fBluetooth where fBluetooth = 2.4 GHz), EM traces were captured in an office corridor up to 15 m. The model was trained on 500K traces from five nRF52 devices across five settings (cable, 1 m, 2 m, 4 m, 8 m) equally, each trace averaging 100 repeated encryptions. Four other nRF52 AES implementations were then attacked at 3, 6, 9, and 15 m (Table IV): the CNN-based attack consistently recovered the partial key with key enumeration complexity 20 and no subkey ranking needed, effective up to 15 m, with trace counts generally increasing with distance (437 traces at 3 m up to 2,475 at 9 m), while the 15 m corridor result needed only 367 traces (×1K repetitions) thanks to clear line-of-sight. The operational constraint in [38], however, is the target performing 1K repeated encryptions per trace — highly impractical remotely without physical access — and the frequent infeasibility of obtaining plaintext, making the first-round AES SBox operation unsuitable as an attack point.

To eliminate the trace-repetition requirement, [85] proposed a noise-aware augmentation profiling strategy: instead of building profiling models from far-field traces whose key-dependent features are too weak, the CNN classifier is trained on high-quality "clean" traces acquired through a coaxial cable connection with manually injected Gaussian noise simulating realistic distortions (Fig. 7), learning better leakage representations. The model achieved a 20-fold efficiency improvement (Table V): recovering a last-round AES sub-roundkey with only 341 traces captured at 15 m in a corridor, without any encryption repetition, using the last-round SBox output as the attack point to remove the known-plaintext obstacle. The master thesis [86] investigated whether more realistic environment-dependent noise augmentation further improves generalization of deep-learning-based FEM-SCA, showing realistic noise augmentation does not always guarantee improved generalization over Gaussian noise, with signal amplitude, trace alignment, and propagation conditions playing more critical roles. In another direction, Zhang et al. [87] introduced hardware assistance with a repeater-assisted FEM-SCA framework placing an intermediate RF repeater between target and receiver to amplify leakage and improve SNR, reducing traces for successful CPA (Table VI: 2,450 traces at 0.2 m and 6,450 at 0.7 m) at the cost of covertness, since the repeater must be deployed close to the victim device. Their subsequent work [88] proposed a spectral addition technique: synchronously capturing screaming-channel signals from two frequency bands (2.464 GHz and 2.528 GHz leaked from the same nRF52832 Bluetooth chip), phase-aligning and coherently summing them to suppress uncorrelated Gaussian noise and enhance SNR, with a propagation model theoretically predicting the SNR gain (Table VII: 3,600 CPA traces at 0.5 m), noting the 6 dBi antennas in [87], [88] versus 24 dBi directional antennas in earlier studies partly explain lower efficiency and shorter distances. In 2023, [89] experimentally demonstrated that leakage-RF coupling is a complex process shaped by substrate, power, and ground-plane propagation, with information-bearing leakage observable at non-harmonic frequencies: scanning 150 wired frequencies from 2.45 GHz to 2.6 GHz in 1 MHz steps then conducting profiled correlation attacks [91] wirelessly up to 7 m, 5 selected non-harmonic frequencies remained highly effective (PGE ≤ 2^32), the best at 2.484 GHz requiring only 65×50 traces for subkey recovery (Table VIII), significantly expanding the attack surface to bypass polluted harmonic bands.

In 2025, advanced deep-learning techniques became a powerful ally: [90] proposed a Transformer [93] based neural network, the Switch-T model, learning sensitive features from different side channels with Elastic Weight Consolidation (EWC) [94] for continual cross-channel profiling (e.g., power consumption) while retaining far-field EM capability against the nRF52 Bluetooth AES implementation, needing only 1 epoch to readjust to the nRF task after profiling on other devices (STM32F3 MCU, ATXmega128D4) and recording 2,976 traces at 15 m (Table IX). Then Wang et al. [92] achieved a leap with a Domain-Adversarial ReFeature Neural Network (DAR-NN) reconstructing key-dependent features from "noisy" traces via adaptive feature-level domain alignment, derived from the Domain-Adversarial Neural Network (DANN) [95] by integrating a Denoising Autoencoder (DAE) with two classification regularizers: on identical equipment as [38], [85] but at center receiving frequency 2.272 GHz (Bluetooth radio 2.4 GHz offset by twice the CPU clock 64 MHz), it recovered a subkey at 15 m with only 141 traces without repetition (a reported 58.7% improvement over prior SOTA), retained robustness under passive and active HackRF interference (771–852 traces) where baseline models [85], [90], [96]–[98] failed with 5K traces, and transferred to attacking STM32 and ATXmega MCUs with power traces (Table X, Fig. 8). In 2026, [99] introduced the multi-screaming-channel attack framework in IEEE-TIFS: treating leakage at different RF frequencies as independent side-channel sources rather than redundant observations, and comparing data fusion, feature fusion (with PCA [100] or LDA [101]), and decision fusion; decision fusion won by analyzing each frequency independently and aggregating key-probability scores: at 15 m two non-harmonic channels reduced traces to reach a target GE from 125×500 and 146×500 individually to only 45×500 with decision fusion, and achieved the first successful 30 m attack in an RF-polluted environment with 2–2.77× fewer traces (Table XI, Fig. 9). Table XII summarizes the chronological progression on nRF52 TinyAES across four stages (Fig. 10): remote feasibility testing, DL-driven enhancement, signal-acquisition enhancement, and robust practicalization — from laboratory curiosity to a practical remote attack framework.

VI. Attack Targets and Variants

Beyond improving attack efficiency on the TinyAES reference target, recent FEM-SCA research has increasingly explored the generality of this leakage mechanism across cryptographic implementations and application scenarios: from TinyAES-based demonstrations to hardware accelerators, post-quantum cryptographic primitives, and practical wireless protocols, revealing broad applicability of RF-mediated side-channel leakage. Note that some included works do not demonstrate long-distance over-the-air attacks, but they still exploit the core screaming-channel leakage mechanism — unintended coupling between digital computation and analog/RF circuitry through which sensitive information is modulated onto RF signals and becomes externally observable — so they remain relevant to the threat model even with cabled or only partially remote setups. In contrast, other remote side-channel works such as the RF-based trapped-ion quantum-computer attack [103], TEMPEST attacks [43], [104], [105], and EM eye [106] do not rely on digital-to-analog/RF coupling in mixed-signal chips; their leakage originates from EM emanations themselves rather than computation-induced leakage coupled into a communication carrier, so they are broader EM studies not detailed here; likewise recent EM eavesdropping attacks [107], [108] exploiting coupling between two EM signals fall outside the RF range and are excluded. In 2022, Ruize et al. explored RF-based leakage from two PQC implementations, CRYSTALS-Kyber and Saber [109], using the screaming channel as the side channel [110], [111]; however, the side-channel signal was captured via a direct conducted SMA coaxial cable connection, not a true remote far-field scenario, since PQC RF leakage is still insufficiently exploitable over the air with current profiling methods — even with cleaner, stronger cable traces attack efficiency remained relatively low, so practical remote extension requires advances in leakage enhancement, acquisition, and robust modeling.

In addition to unprotected AES, [26] conducted the first screaming channel attack on a Bluetooth device implementation of masked AES in 2024, targeting the Rivain–Prouff (RP) masking scheme [113]: since TinyAES is seldom deployed where computational resources are not severely constrained and security-critical applications typically use countermeasures [114], they evaluated to what extent RP-masked AES can be compromised under FEM-SCA. Keeping the hardware setup as in [38], [85] except the receiving antenna, the attack focused on MixColumns leakage with results (Table XIII: 290K traces ×1 at 0.1 m with enumeration 20 and 500K×100 profiling) showing inefficiency from two causes: RP masking is provably secure against lower-order attacks by splitting sensitive variables into random shares, forcing complex higher-order analysis, and the already faint signal was further degraded by replacing the TL-ANT 2424B grid parabolic antenna (24 dBi gain) with a VERT2450 vertical antenna (3 dBi gain), yielding very low SNR. The same group in [115] then improved efficiency by 80% by shifting the attack point to the ShiftRow operation, using 50K traces without repetition for partial-key recovery from RP-masked AES. While all reviewed FEM-SCAs focused on software AES, [112] demonstrated in 2025 the first successful screaming channel attack on a hardware AES accelerator: exploiting unintentional RF leakage from the crypto part during AES-CCM mode encryption, targeting the hardware accelerator of the same nRF52832 Bluetooth SoC victim as in [38], [85], [92]. Since hardware crypto implementations are significantly more side-channel resilient due to parallelism, pipelining, and synchronous clocking [116], [117], the core setup used a USRP-2944R SDR receiver demodulating RF at center frequency 2.528 GHz (Bluetooth channel 2.4 GHz offset by eight times the AES accelerator clock 16 MHz); leakage analysis on a known-key reference device inferred proprietary accelerator internals (apparently fully parallel AES rounds, state register closely related to last-round SBox input), targeting encryption of counter block Ctr1 in CCM mode derivable from known message and ciphertext. The key-recovery stage combined correlation analysis with key enumeration: each trace averaged 10K repeated AES-CCM encryptions to improve SNR, and with the Hamming weight of the last-round SBox input as leakage model, correlation recovered most last-round key bytes — four of sixteen round-key bytes misranked for antenna traces at 1 m — and as Table XIV shows, 180K traces at 1 m recovered the full key with enumeration complexity approximately 2^42.8: the hardware AES-CCM accelerator is not immune, but attacking it via FEM-SCA remains substantially harder than software TinyAES. Building on this feasibility study, [118] improved efficiency with ensemble learning: training 3×16 MLP models (three per byte position for ensemble voting) on 300K cable traces (each averaged from 10K encryptions) using five short intervals around last rounds, then attacking two unseen devices via cable or 1 m antenna; with 90K traces at 1 m (each again averaged from 10K encryptions) the full AES key was recovered after key enumeration with complexity 2^24 (Table XV) — profiled learning makes hardware-accelerator FEM-SCA more efficient, but still requires massive repeated encryptions, a large cable-based profiling set, and carefully controlled acquisition.

Shifting focus from cryptographic key extraction, [25] demonstrated that a wide range of digital components within a modern embedded system can unintentionally leak raw sensitive information through the device's own radio transmitter: signals from internal buses, serial ports, memory accesses, and even external protocols like NFC can be modulated and amplified by the RF front-end, enabling direct far-field readout. Using the same SDRs and antenna as in [23], [24], [38], [85] on common IoT devices (Nordic nRF, ESP32, TI CC2642), an adversary successfully demodulated and extracted clear-text data such as UART bytes or memory values from a single trace at distances up to 20 m; even without galvanic connection, signals from components tens of centimeters away on the PCB can be picked up, amplified, and broadcast, proving the pervasiveness and severity of this channel. If an attacker can inject malicious code (via malware or a software vulnerability), they are no longer limited to incidental leakage but can orchestrate it: injected code actively modulates sensitive data in a way perfectly tailored for long-range RF exfiltration. Attack specifications for UART byte-value classification trained a Long Short-Term Memory (LSTM) model on only a few hundred 16-sample traces to extract "clear-text" data (0x00 to 0x0f) with 60% single-trace accuracy (Table XVI), with distance inferred up to 20 m as their general setup's maximum far-field range. Another extension toward satellite-related hardware in [119] studied several small-satellite-relevant sub-GHz SoC transceivers for processor-activity-correlated RF leakage: following the conceptual discussion in [120] (leakage visible when continuous RF transmission changes into a periodic dashed or intensity-varying pattern synchronized with the processor sleep-encryption cycle), the TI CC1111 SoC showed a clear dashed-line pattern around its intended band (potential screaming-channel leakage path), the TI CC1310 weaker evidence (amplitude changes correlated with processor state), and the Seeed Wio-E5 LoRa-like SoC no clear relationship — a useful qualitative leakage-screening study without quantitative FEM-SCA results. More recently, ModelSpy, a DNN remote attack framework in NDSS'26 [121], extended far-field EM analysis beyond cryptographic implementations and RF-enabled embedded SoCs to complex computing workloads: rather than secret keys, it targets black-box DNN architecture snooping from GPUs, reconstructing the number and types of neural-network layers and layer-wise hyperparameters — knowledge improving subsequent black-box attacks [122]–[125]. Unlike conventional screaming channels, ModelSpy relies on no intentional RF carrier from a wireless module; the carrier originates from continuous radiative EM emissions of GPU hardware (voltage regulators, memory clocks, memory-refresh circuits [126]) at MHz-to-GHz frequencies serving as high-frequency carriers. Experiments with a USRP B210 receiver and omnidirectional or 20×20 cm panel antenna around 5 GHz center frequency at 8 MHz sampling rate, on five consumer NVIDIA GPUs and diverse DNN architectures (AlexNet [127], VGG [128], ResNet [129], BERT [130], plus randomly generated CNN, LSTM, Transformer, and linear architectures), achieved 97.6% average layer-segmentation accuracy and 94.0% hyperparameter-estimation accuracy on unseen architectures, effective up to approximately 6 m (dropping to 42.3% at 7 m) — far beyond millimeter-range techniques near the GPU or its power cable [5], [131]. Through-wall attacks through wood, concrete, and glass still achieved 96.0%/87.7%, 96.1%/85.7%, and 96.3%/87.3%, an active WiFi access point at 5.18 GHz only 1 m from the victim degraded accuracies by 8.7/7.7 percentage points, and a larger parabolic antenna extended range beyond 10 m (Table XVII) — a notable shift from cryptographic secret recovery toward high-level semantic inference of complex workloads.

VII. Public FEM-SCA Data Collections

Public FEM-SCA data collections are essential for transforming screaming-channel research from isolated proof-of-concept demonstrations into a reproducible and comparable field: unlike power or near-field EM analysis, far-field leakage is strongly affected by propagation and environmental interference, and without shared collections it is difficult to tell whether an improvement comes from a genuinely stronger attack model or a cleaner measurement setup. To date, six public collections have provided reusable experimental foundations (Fig. 2): the original CCS'18 collection [23], the CHES'20 collection [24], the ASHES'20 DL-oriented collection (first-round AES) [38], the CPSS'21 DL-oriented collection (last-round AES) [85], the ACSAC'24 BlueScream collection [80], and the NDSS'26 GPU-DNN collection [121]. These resources are not yet organized or systematic enough to be regarded as standardized datasets such as [132]–[134], often covering only isolated distances, devices, or scenarios without unified structure and labeling format, so they are termed collections rather than datasets. Note that some works claiming released data are not counted as independent collections: [112], [118] state data and scripts are available via a KTH BOX link currently redirecting to an SSO login page inaccessible outside KTH, and although [92] states code and traces are public, its evaluation traces derive from the CPSS'21 collection [85]. The first collection [23] established fundamental feasibility of remote EM leakage exploitation: released traces stored as NumPy arrays in three subsets (a tinyaes_anechoic_10m_attack subset with 3,000 attack traces, a tinyaes_anechoic_10m_template subset with 243,911 template/profiling traces, and an mbedtls_1m_home subset with 38,645 traces), each with a fixed AES key file and plaintext file linking traces to known cryptographic inputs, in a ~2.6 GB repository (Table XVIII), with each trace averaging 500 measurements of the same repeated encryption. For the CCS'18 collection we further examined first-round SBox-output leakage via TVLA [136] on the first key byte with a fixed-versus-random partition on the intermediate Hamming weight: on the tinyaes_anechoic_10m_template subset, 80K traces split by HW(SBox(P[0]⊕K[0])) > 4 versus < 4 (excluding HW = 4) yielded Welch's t-test leakage around the later trace part with maximum absolute t-value 8.659 at sample index 848 (exceeding the common |t| > 4.5 threshold [36]) within first-round AES after the initial AddRoundKey stage, from which the remaining execution structure was visually derived (Fig. 12a: pre-activities like key expansion, initial AddRoundKey, first-round SubBytes, ShiftRows, MixColumns, AddRoundKey, then second-round activity). For the mbedTLS subset among 38,645 traces the t-test showed clear leakage around sample index 375 with maximum absolute t-value 13.958 and five samples exceeding the threshold (Fig. 12b).

The CHES'20 collection [24] extended the CCS'18 proof-of-concept traces into a systematic resource for physical and statistical leakage properties: while CCS'18 primarily captured AES leakage through quadrature amplitude demodulation centered at 2.528 GHz, screaming-channel leakage from the same Nordic nRF52 Bluetooth 5 target in the 2.4 GHz band is not confined to a single frequency but emerges at multiple harmonics, so CHES'20 collected traces at multiple leakage frequencies for systematic frequency-dependent examination (Table XIX: 119,792 traces via compare_frequencies with 36,000 TinyAES traces at 10 cm for four frequencies 2.528, 2.530, 2.554, and 2.608 GHz with variable-key profiling and fixed-key attack traces, hwcrypto with 75,602 traces from the dedicated AES-128 hardware block where traces reveal plaintext/key/ciphertext transfers between memory and the hardware block but internal round operations are not clearly visible — showing hardware AES is substantially harder than software TinyAES — and shannon_080719 with 16,190 traces capturing a longer full-encryption window supporting full-trace characterization and round-level localization). Although the GitHub page claims coverage from leakage characterization to 34 m detection, detailed inspection of the downloaded package found only these three organized subsets. The ASHES'20 DL-oriented collection (first-round AES) [38] then shifted focus toward attack scalability under systematically varied remote conditions: traces collected across gradually changing distances on a Nordic nRF52 Bluetooth SoC running software TinyAES AES with an Ettus USRP N210 SDR and TP-Link TL-ANT2424B parabolic antenna, in for_training and for_testing subsets of 400-sample first-round traces (training at cable, 1 m, 2 m, 4 m, 8 m; testing at 3 m, 6 m, 9 m, 15 m with 1, 100, and 1000 encryption-repetition variants) totaling 1,520K traces (Table XX) — an intermediate step between leakage-understanding releases and deep-learning testbeds. Following it, the CPSS'21 collection [85] refined data design by separating profiling and attack conditions aggressively: cleaner cable profiling traces (100K traces ×100 averaged from five profiling devices) then generalization to 15 m far-field traces from victim devices (75K traces with 1, 10, 100 averaged variants) on the same nRF52832 TinyAES target with an N210 receiver tuned to 2.528 GHz at 5 MHz sampling rate, totaling 275K traces (Table XXI) — shifting dataset focus from distance-dependent generalization to cross-domain learning between high-SNR cable traces and low-SNR far-field traces. Figure 14 shows the close relationship between the collections (400-sample window, similar amplitude-based waveform representation, operation-level temporal segmentation), differing in AES window: ASHES'20 covers execution start (pre-activities, initial AddRoundKey, first-round operations) while CPSS'21 covers execution end (part of ninth round plus last round: SubBytes, ShiftRows, AddRoundKey). CPSS'21 targets the last-round key rather than the original cipher key directly: the adversary learns last-round intermediate leakage then derives the original AES key through the inverse key schedule, so every last-round key byte must be recovered correctly or reconstruction fails or enumeration cost soars — motivated by the realistic remote assumption that plaintext may be uncontrollable or unobservable while ciphertext is accessible from transmitted data or protocol output, so last-round intermediates are computed from known ciphertext and guessed last-round key bytes.

In 2024, the BlueScream data collection was built and released in [80] at ACSAC'24, shifting from controlled-firmware TinyAES traces toward a realistic BLE protocol-stack scenario: targeting the BLE key derivation (re-keying) mechanism on a Nordic nRF52832 development kit running NimBLE-based firmware [137], with three firmware variants (Fcustom for controlled experiments, Finstru for instrumented BLE evaluation with controllable AES inputs, and Fdefault for a stock BLE setting where only one AES input is known); by manipulating BLE communication the attacker forces the victim to transmit during the cryptographic operation so the RF front-end broadcasts the leakage. Publicly available traces correspond mostly to the most realistic Fdefault setting in two subsets (pairing_10cm_anechoic: ~65K profiling and 16K testing traces at 10 cm in an anechoic chamber with an omnidirectional WiFi antenna; 2lna_highgain_norep: ~65K profiling and 41K testing traces at 1.2 m in an office with a high-gain directional antenna plus two LNAs) totaling 187,971 traces of 4,000 samples each (Table XXII) — well suited for realistic wireless-protocol attack scenarios though centered on BLE AES leakage. In 2026, the NDSS'26 GPU-DNN Collection [121] extended public far-field data from cryptographic implementations to GPU-based DNN inference workloads: 20,063 traces of 5,023 samples each across three subsets (cross_GPU_evaluation: 10,696 traces from five NVIDIA GPUs at 1 m with representative CNN, LSTM, and Transformer inference workloads; distance_evaluation: 6,946 traces from an RTX 3060 Ti in the open field from 0 to 7 m; cross_wall_evaluation: 2,421 traces through concrete, glass, and wood walls) (Table XXIII) — whose major distinction is shifting the leakage target from cryptographic implementations to high-level DNN architecture information. The six collections form a staged route (Fig. 15): the CCS'18 and CHES'20 collections from the EU-RECOM group [23], [24] establish basic feasibility and physical understanding of RF-mediated AES leakage on mixed-signal Bluetooth SoCs (TinyAES, mbedTLS [135], and hardware AES under anechoic-chamber and office environments at 0.1 m, 1 m, and 10 m); the ASHES'20 and CPSS'21 DL-oriented collections from KTH [38], [85] evaluate whether data-driven models can exploit such leakage (first-round versus last-round recovery assumptions, distance-induced distribution shifts, model transferability); and the ACSAC'24 BlueScream and NDSS'26 GPU-DNN collections [80], [121] move toward realistic and diverse targets (BLE re-keying AES leakage requiring interaction with BLE communication, and far-field DNN inference leakage across five GPU models). The six public collections thus form a chronological experimental roadmap rather than a uniform benchmark suite: proof-of-concept and mechanism-understanding foundations, deep-learning-oriented trace resources, and practical-target extensions.

VIII. Discussion: How to Deal with the Radio

The preceding sections traced FEM-SCA development from discovery and physical understanding of screaming-channel leakage, through enhanced attack methodologies and target variants, to public trace collections. Across these studies a consistent pattern emerged: improvements in one dimension often expose constraints in another, such as longer attack distance requiring more traces, stronger models depending on narrower target assumptions, or more realistic protocol scenarios leading to lower attack efficiency. This section distills the discussion into one key trade-off, three major limitations, and three future research roadmaps. The key trade-off lies between remote attack capability and attack efficiency: once screaming-channel attacks acquire far-field exploitation ability, the most direct cost is a substantial reduction in attack efficiency compared with conventional power and near-field EM attacks, because a longer-range capability is obtained by inserting an additional physical path between the sensitive computation and the measurement device. In power analysis, leakage is measured from power consumption caused directly by data-dependent switching; in near-field EM analysis, from localized radiation near the circuit or PCB; in FEM-SCA, by contrast, digital leakage must first couple into the analog/RF transmission chain, modulate or distort the radio signal, propagate through space, and then be captured by the receiving antenna. Each step makes the leakage more accessible from a distance but weakens, mixes, and distorts the key-dependent information. The far-field nature of FEM-SCA is therefore not a free improvement over traditional SCAs: it expands the attack surface spatially but sharply reduces the signal quality available for statistical or learning-based exploitation. This efficiency loss is visible comparing similar cryptographic targets: for unprotected AES implementations on ARM-class microcontrollers, direct deep-learning-based power-analysis attacks can recover the full key with fewer than ten traces in some reported settings [138], [139] because the power trace contains strong temporally localized data-dependent leakage, whereas in a screaming-channel setting the same sensitive operation is observed as an RF waveform in which useful leakage has been coupled into a much stronger communication signal and contaminated by channel effects, receiver noise, and unrelated radio activity — so even at very short receiver distances FEM-SCA typically requires several hundred traces to recover a subkey [85]. This difference is not merely immature attack algorithms but a basic physical constraint: the coupling that makes leakage radiate also reduces its amplitude, separability, and stability, so the attack model must learn from weaker, less repeatable features entangled with the RF carrier and environment. The second layer of this trade-off appears as attack distance increases: the RF signal suffers path loss, environmental attenuation, multipath propagation, misalignment, and interference from other transmitters, reducing the effective SNR of the leakage component even when the radio communication signal itself can still be received — an important distinction: a receiver may observe the radio transmission while the embedded side-channel component remains too weak for key recovery or even segment alignment. Increasing distance therefore usually forces compensation: more traces, averaged repeated encryptions, higher-gain directional antennas, low-noise amplifiers, cleaner cable profiling, multiple receiving channels, or more complex neural networks. Each compensation increases attack cost; for example, repeated encryption improves averaged leakage quality but requires observing many identical cryptographic executions, highly impractical in most scenarios. The distance–efficiency relationship should not, however, be described as simple linear or exponential decay: RF propagation is spatially irregular, and reflections from walls, floors, objects, furniture, and human bodies create constructive and destructive interference, so a farther point may receive a stronger RF component than a closer one near a constructive multipath region, while a nearby receiver may suffer fading or poor antenna orientation. Attack efficiency may therefore decrease in a spiral-like, non-monotonic manner: the overall trend worsens with distance but local conditions temporarily improve or degrade leakage — which also explains why distance alone is insufficient to characterize an FEM-SCA experiment, since the same distance can correspond to very different attack difficulty depending on room geometry, antenna placement, obstacles, receiver orientation, carrier frequency, and radio traffic. This non-monotonicity makes trace collection itself a core technical challenge: in traditional power analysis, once probe and trigger are stable, collecting thousands of traces is straightforward, whereas each FEM-SCA trace depends not only on cryptographic execution but on wireless transmission state, RF channel condition, synchronization quality, and environmental stability. The attacker may need to wait for suitable protocol events, trigger repeated pairing or encryption operations, align traces captured through noisy RF reception, and discard corrupted measurements; with frequency hopping, packet scheduling, or variable radio activity, acquisition becomes extremely hard, and collecting hundreds or thousands of usable traces from a remote victim may require prolonged observation and carefully positioned equipment. From the authors' perspective, this trade-off should become a central evaluation principle: reporting only that an attack succeeds at a certain distance can be misleading since distance does not capture full attack cost, and similarly reporting only trace counts can hide that traces required a far larger set of captured but uninvolved data [92]. A more meaningful evaluation should jointly report attack distance, number of attack traces, number of captures to get a useful trace, quantitative environmental conditions, and synchronization effort — only then can the field distinguish technically feasible attacks under strong laboratory assumptions from practical remote threats.

Although recent FEM-SCA studies have made clear progress in attack distance, trace efficiency, modeling capability, and protocol awareness, the field is still shaped by structural limitations that do not invalidate existing results but define the boundary of what can currently be concluded: the evidence base remains concentrated on a narrow set of platforms, a narrow set of wireless technologies, and a fragmented set of public trace releases, making it difficult to assess whether FEM-SCA is a broadly generalizable threat or a set of strong but platform-dependent demonstrations. The first and most visible limitation is limited platform generality: concentration on the Nordic nRF52 family [23], [24], [38], [85], [99] made cross-paper comparison easier, supported public-trace reuse, and provided a stable reference platform for studying digital-to-RF coupling — nRF52 played a benchmark-device role for early-stage FEM-SCA research — but the same concentration limits external validity. FEM-SCA leakage is not determined only by the cryptographic algorithm or software implementation but by a cross-layer physical path involving the processor, memory bus, clock tree, power distribution network, analog/RF front-end, PCB layout, antenna path, firmware scheduling, and radio transmission behavior; changing any layer may alter leakage strength, stability, and frequency-domain structure. Evidence from masked AES, hardware AES accelerators, and BLE protocol-level implementations already shows target or context changes significantly reducing attack efficiency [26], [115], [118], so a model or acquisition strategy working well on one nRF52 board cannot be assumed to transfer directly to other wireless SoCs even with the same cryptographic primitive. The authors' view is that the next stage should treat cross-platform validation as a core requirement rather than an optional extension, without which current conclusions about FEM-SCA practicality remain persuasive for a well-studied device family but uncertain for the broader embedded wireless range. The second limitation is that quantitative evidence remains overwhelmingly Bluetooth-centric: foundational screaming-channel studies exploit leakage coupled into Bluetooth transmissions and major public collections rely predominantly on Bluetooth-related acquisition settings [23], [24], [38], [80], [85], leaving the broader radio landscape underexplored. WiFi and Zigbee have different carrier structures, bandwidths, modulation schemes, packet durations, channel-access mechanisms, and baseband processing pipelines, which may change not only how leakage couples into the RF signal but also how traces should be synchronized, filtered, segmented, and modeled. While some works suggest digital activity can be picked up and radiated by wireless front-ends beyond a single cryptographic benchmark [23], [37], systematic quantitative evidence for WiFi- or Zigbee-based FEM-SCA is still lacking — key-recovery success rate, trace complexity, attack distance, profiling requirement, and robustness under realistic conditions. The authors' view is that other radio channels such as WiFi and Zigbee should not be treated as simple extensions of Bluetooth-based FEM-SCA where a receiver-frequency change alone captures the full difference; instead each wireless protocol should be studied as a distinct leakage environment with its own modulation, timing, and interference structure — necessary evidence before FEM-SCA can be described as a general RF-mediated side-channel threat rather than one currently supported mainly by Bluetooth experiments. The third limitation is the absence of a structured FEM-SCA benchmark covering diverse attack conditions under the same experimental setting: existing public collections were built for different purposes and acquisition assumptions — CCS'18 and CHES'20 mainly supporting proof-of-concept demonstration [23], [24], ASHES'20 and CPSS'21 supporting deep-learning-based evaluation on first- and last-round TinyAES leakage [38], [85], and BlueScream moving toward BLE protocol-level leakage [80] — so they do not form a unified benchmark where different algorithms, platforms, distances, environments, and obstacles are evaluated under a consistent experimental configuration. This makes cross-paper and cross-method comparison difficult and, more importantly, provides no controlled variable matrix: the field still lacks a public benchmark measuring multiple cryptographic algorithms, multiple wireless SoCs, multiple distances, indoor and outdoor environments, and different obstacle conditions while keeping receiver, antenna, sampling rate, synchronization method, firmware design, and data format consistent. Without such a benchmark it is difficult to determine whether a reported improvement comes from a genuinely better attack method or a more favorable acquisition condition: a model appearing more efficient on one collection may simply benefit from less environmental distortion and better trace alignment. The authors' view is that the field needs a structural FEM-SCA benchmark designed around controlled comparison rather than isolated reproduction, with a systematic experimental grid — its absence is not merely a data-sharing problem but a methodological bottleneck limiting the maturity and comparability of FEM-SCA research.

The limitations above suggest the next stage of FEM-SCA research should not focus only on improving attack success rates under already familiar settings; three research roadmaps may move the field from isolated demonstrations toward mature security evaluation. First, the systematic design and evaluation of countermeasures: although existing FEM-SCA works have mainly focused on attack construction, the demonstrated leakage mechanism already suggests the need for protection in security-critical wireless devices. Since FEM-SCA is generated through coupling between digital computation and the analog/RF transmission chain, countermeasures should not be limited to traditional cryptographic masking or hiding [140]. Effective side-channel countermeasure design generally follows two core principles [141]: (1) degenerating the statistical relationship between secret data and physical leakage, and (2) increasing the effort for an adversary to capture exploitable traces. Applying these principles to screaming channels suggests three levels. At the Algorithmic Level, the focus is modifying the cryptographic computation itself to degenerate or remove data-dependency of leakage, directly addressing the first principle by breaking the statistical link between secrets and emanations — most prominently masking [142], [143], which splits every sensitive variable into multiple random shares and performs computation on these shares, so EM emanation at any moment correlates with a random value rather than the actual secret; for screaming channels relying on distinct EM signatures of specific data values, effective masking renders the EM radiation profile statistically random, removing the exploitable signal. At the Implementation Level, countermeasure design changes device operation in traces to disrupt trace alignment and analysis, serving the second principle by greatly increasing the traces an adversary must collect and processing complexity — techniques such as clock jitter [144], [145], duplicating cryptographic execution [146], [147], or inserting dummy operations create desynchronization between internal device state and measured traces [148], [149]: even with data-dependent leakage present, desynchronization prevents perfect alignment of thousands of traces, so leakage SNR stays low and required measurements rise substantially [150], often impractical for remote attacks where trace quality is already limited. At the Physical Level, directly increasing adversary effort: a highly effective strategy is actively introducing interference to pollute the electromagnetic spectrum, as tested in [92]; beyond power-consuming active interference sources, shielding — enclosing the device or sensitive components within a conductive cage — is another primary avenue against screaming channel attacks, with different cover types showing varying shielding effectiveness against near-field EM leakage [151]–[153]; the significant designer challenge is the inherent trade-off between side-channel resistance and intended RF communication performance, since a shield blocking leakage can also degrade legitimate wireless signals [154]. Second, different attack targets and RF protocols: future studies should evaluate different boards, SoC vendors, PCB layouts, and cryptographic implementations — not only software implementations but hardware accelerators, masked implementations, post-quantum primitives, authenticated encryption, and protocol-level security operations — and move beyond Bluetooth-enabled SoCs to examine WiFi, Zigbee, and other wireless protocols with distinct RF architectures, operating frequencies, modulation schemes, and transmission behaviors that may significantly affect coupling and propagation of secret-dependent leakage. Future studies should investigate whether exploitable leakage can be consistently observed across these protocols and, more importantly, demonstrate practical vulnerability through quantitative key-recovery results under realistic remote attack settings; cross-target and cross-protocol studies are essential to determine whether FEM-SCA is a general attack class or a threat currently concentrated on several favorable platforms and cryptos, requiring comparative experiments with consistent acquisition and reporting standards. Third, boundary analysis and modeling: existing studies show FEM-SCA works under certain distances, environments, and target assumptions, but the field still lacks a clear model of where attacks become infeasible. Future work should therefore characterize FEM-SCA boundary conditions, including maximum useful distance, minimum required leakage strength, tolerable interference level, required repetition count, antenna placement sensitivity, obstacle effects, and protocol-triggering constraints. This requires both empirical measurement and theoretical modeling: empirical measurements map where leakage remains distinguishable and where it collapses into noise, while analytical models relate RF propagation and digital-to-RF coupling to side-channel success indicators such as key rank or classification confidence. Such modeling would make FEM-SCA evaluation more predictive: instead of only reporting that an attack succeeds in one setting, researchers could estimate conditions under which the same attack likely remains effective — helping defenders identify high-risk deployment scenarios and design meaningful security tests for wireless embedded devices.

IX. Conclusion

This paper provides the first systematic view of screaming channel attacks, or FEM-SCAs, as a distinct class of radio-coupling-based remote side-channel attacks. By organizing existing studies around leakage understanding, attack enhancement, target variants, and public data collections, we establish a coherent picture of how this research area has evolved since the first demonstration of the screaming-channel phenomenon. The reviewed works show that FEM-SCA is no longer only a lab curiosity: sensitive computation can be carried by ordinary RF transmissions and recovered at a distance under certain conditions. Early studies primarily focused on establishing the existence and physical origin of secret-dependent leakage unintentionally coupled into legitimate RF transmissions, whereas subsequent research progressively improved attack efficiency, robustness, and operating distance — from the first template attacks at 10 m [23] to multi-distance-device CNN attacks [38], noise-augmented cable profiling with 341 traces at 15 m [85], adversarial feature reconstruction with 141 traces [92], and successful multi-channel fusion at 30 m [99] — and more recent efforts have further explored different implementation settings and increasingly realistic wireless environments, gradually shifting FEM-SCA research from proof-of-concept demonstrations toward practical remote attack scenarios. These developments establish FEM-SCA as an important extension of conventional side-channel analysis and highlight a previously underestimated attack surface in RF-enabled embedded and IoT systems.

At the same time, the available evidence also defines a clear boundary for the current state of FEM-SCA research. Existing attacks still face an efficiency–distance trade-off, as increasing the measurement distance typically weakens the exploitable leakage and substantially increases the effort required for successful analysis. Moreover, current studies remain largely concentrated on Bluetooth-enabled platforms, while the feasibility and characteristics of such leakage under other RF protocols and device architectures have received only limited investigation. Although several public datasets have been released to improve reproducibility, they are generally collected under different experimental conditions and therefore do not yet constitute a structured benchmark with unified acquisition settings and evaluation criteria. This lack of standardized evaluation makes direct and fair comparison between existing attacks challenging and leaves uncertainty about whether techniques demonstrated under specific experimental setups can generalize to different cases. Addressing these limitations is therefore essential for moving FEM-SCA research from isolated demonstrations toward systematic and practically relevant security evaluation. Future FEM-SCA research therefore needs to become more comparative, more cross-platform, and more countermeasure-aware. Progress should be measured not only by whether an attack succeeds, but also by the cost, assumptions, distance, environment, and target conditions under which it succeeds. A mature understanding of screaming channel leakage may require both stronger attacks and clearer failure boundaries.

References

• [18] P. Kocher, J. Jaffe, and B. Jun, “Differential power analysis,” in Annual International Cryptology Conference. Springer, 1999, pp. 388–397.
• [23] G. Camurati, S. Poeplau, M. Muench, T. Hayes, and A. Francillon, “Screaming channels: When electromagnetic side channels meet radio transceivers,” in ACM CCS, 2018, pp. 163–177.
• [24] G. Camurati, A. Francillon, and F.-X. Standaert, “Understanding screaming channels,” TCHES, pp. 358–401, 2020.
• [25] M. Guri, “Air-gap electromagnetic covert channel,” IEEE TDSC, vol. 21, no. 4, pp. 2127–2144, 2023.
• [26] H. Wang, “Amplitude-modulated EM side-channel attack on provably secure masked AES,” Journal of Cryptographic Engineering, pp. 1–13, 2024.
• [34] S. Chari, J. R. Rao, and P. Rohatgi, “Template attacks,” in CHES. Springer, 2002, pp. 13–28.
• [35] E. Brier, C. Clavier, and F. Olivier, “Correlation power analysis with a leakage model,” in CHES. Springer, 2004, pp. 16–29.
• [36] B. J. Gilbert Goodwill et al., “A testing methodology for side-channel resistance validation,” in NIST Non-Invasive Attack Testing Workshop, 2011.
• [37] D. Lei et al., “A review of security analysis technologies for screaming channel based on RFSoC,” in ICIBA. IEEE, 2026, pp. 1–11.
• [38] R. Wang, H. Wang, and E. Dubrova, “Far field EM side-channel attack on AES using deep learning,” in ASHES. ACM, 2020, pp. 35–44.
• [48] H. Wang et al., “How diversity affects deep-learning side-channel attacks,” in NORCAS. IEEE, 2019, pp. 1–7.
• [64] E. Cagli, C. Dumas, and E. Prouff, “Convolutional neural networks with data augmentation against jitter-based countermeasures,” in CHES. Springer, 2017, pp. 45–68.
• [67] Y. Ji et al., “A side-channel attack on a hardware implementation of CRYSTALS-Kyber,” in ETS. IEEE, 2023, pp. 1–5.
• [68] G. Zaid et al., “Conditional variational autoencoder based on stochastic attacks,” TCHES, pp. 310–357, 2023.
• [72] K. S. Subramani et al., “ACE: Adaptive channel estimation for detecting analog/RF Trojans in WLAN transceivers,” in ICCAD. IEEE/ACM, 2017, pp. 722–727.
• [75] D. R. Gnad, J. Krautter, and M. B. Tahoori, “Leaky noise: New side-channel attack vectors in mixed-signal IoT devices,” TCHES, pp. 305–339, 2019.
• [78] M. Ettus and M. Braun, “The universal software radio peripheral (USRP) family of low-cost SDRs,” 2015, pp. 3–23.
• [79] “Small portable AES128/192/256 in C,” Github, 2013, https://github.com/kokke/tiny-AES-c/.
• [80] P. Ayoub et al., “BlueScream: Screaming channels on Bluetooth Low Energy,” in ACSAC. IEEE, 2024, pp. 636–649.
• [81] H. Wang, “Deep learning side-channel attacks on AES,” Ph.D. dissertation, KTH, 2023.
• [82] M. Nomura et al., “A case study of side-channel authentication using screaming channel,” in MASS. IEEE, 2024, pp. 488–489.
• [85] R. Wang et al., “Advanced far field EM side-channel attack on AES,” in CPSS. ACM, 2021, pp. 29–39.
• [87] Z. Hong-Yi et al., “Wireless side-channel analysis method based on repeater (in Chinese),” Journal of Cryptologic Research, vol. 9, no. 1, pp. 175–188, 2022.
• [88] Z. Hong-Yi et al., “Wireless side-channel analysis method based on spectral addition (in Chinese),” Journal of Cryptologic Research, vol. 10, no. 4, pp. 862–878, 2023.
• [89] J. Guillaume et al., “Attacking at non-harmonic frequencies in screaming-channel attacks,” in CARDIS. Springer, 2023, pp. 87–106.
• [90] J. Liao et al., “Switch-T: A novel multi-task deep-learning network for cross-device side-channel attack,” JISA, vol. 93, p. 104146, 2025.
• [92] H. Wang et al., “Feature reconstruction: Far field EM side-channel attacks in complex environment,” IEEE TIFS, vol. 20, pp. 10 066–10 081, 2025.
• [93] A. Vaswani et al., “Attention is all you need,” NeurIPS, vol. 30, 2017.
• [95] Y. Ganin et al., “Domain-adversarial training of neural networks,” JMLR, vol. 17, no. 59, pp. 1–35, 2016.
• [99] J. Guillaume et al., “Multi-screaming-channel attacks: Frequency diversity for enhanced attacks,” IEEE TIFS, vol. 21, pp. 4360–4372, 2026.
• [109] M. J. Kannwischer et al., “PQM4: Post-quantum crypto library for the ARM Cortex-M4,” https://github.com/mupq/pqm4.
• [110] R. Wang, K. Ngo, and E. Dubrova, “Side-channel analysis of Saber KEM using amplitude-modulated EM emanations,” in DSD. IEEE, 2022, pp. 488–495.
• [111] R. Wang and E. Dubrova, “A message recovery attack on LWE/LWR-based PKE/KEMs using amplitude-modulated EM emanations,” in ICISC. Springer, 2022, pp. 450–471.
• [112] Y. Ji, E. Dubrova, and R. Wang, “Screaming channels revisited: Encryption key recovery from AES-CCM accelerator,” in ISCAS. IEEE, 2025, pp. 1–5.
• [113] M. Rivain and E. Prouff, “Provably secure higher-order masking of AES,” in CHES. Springer, 2010, pp. 413–427.
• [115] B. Cao et al., “Exploiting ShiftRow vulnerabilities in addition-chain masked AES,” Microelectronics Journal, p. 106866, 2025.
• [118] Y. Ji, E. Dubrova, and R. Wang, “Is your Bluetooth chip leaking secrets via RF signals?” in ISMVL. IEEE, 2025, pp. 141–146.
• [119] T. Gallagher, “Identifying system-on-a-chip data leaks over radio transmissions of small satellites,” Int. J. Applied Technology and Leadership, vol. 3, no. 1, pp. 1–19, 2024.
• [121] R. Xiao et al., “Peering inside the black-box: Long-range and scalable model architecture snooping via GPU electromagnetic side-channel,” in NDSS, 2026.
• [132] R. Benadjila et al., “Deep learning for side-channel analysis and introduction to ASCAD database,” Journal of Cryptographic Engineering, vol. 10, no. 2, pp. 163–188, 2020.
• [135] “Mbed TLS documentation hub,” https://mbed-tls.readthedocs.io/, accessed: 2026-08-18.
• [136] G. T. Becker et al., “Test vector leakage assessment (TVLA) methodology in practice,” in International Cryptographic Module Conference, 2013.
• [137] A. Mynewt, “Bluetooth Low Energy user guide: NimBLE,” https://mynewt.apache.org/, accessed: 2026-08-21.
• [138] D. He et al., “PD-Net: Learning device-invariant representations for heterogeneous cross-device side-channel attacks,” in DAC, 2026.
• [139] A. Liu et al., “CL-SCA: A contrastive learning approach for profiled side-channel analysis,” IEEE TIFS, vol. 20, pp. 5109–5122, 2025.
• [140] H. Groß and S. Mangard, “A unified masking approach,” Journal of Cryptographic Engineering, vol. 8, no. 2, pp. 109–124, 2018.
• [142] B. Gigerl et al., “Smooth passage with the guards: Second-order hardware masking of the AES,” TCHES, vol. 2024, no. 1, pp. 309–335, 2024.
• [144] B. Hettwer et al., “Lightweight side-channel protection using dynamic clock randomization,” in FPL. IEEE, 2020, pp. 200–207.
• [146] M. Moraitis et al., “A side-channel resistant implementation of AES combining clock randomization with duplication,” in ISCAS. IEEE, 2023, pp. 1–5.
• [151] P. Clerico et al., “Assessment of broadband shielding effectiveness of composite panels for protective enclosures,” IEEE TEMC, vol. 64, no. 5, pp. 1750–1757, 2022.
• [154] D. Park, J. Jeon, and H. Kim, “Antenna sensitivity variation depending on the installation of a shield can,” Microwave and Optical Technology Letters, vol. 62, no. 6, pp. 2311–2315, 2020.
• ... [full reference list of 154 entries available in the original paper (paper.pdf)]