Physical Attacks on End-to-End Autoencoder-Based Communications via Evolutionary Transfer Learning
Evolutionary-based Transfer Learning for Physical Attack on End-to-End Autoencoder Communications
The authors propose physical-layer adversarial attacks PAGA (white-box) and PAMFGA (black-box) against end-to-end autoencoder communication systems, and report that their BLER is higher than baselines such as FGMA and RMAEP.
The attacker injects a power-constrained (PSR) universal perturbation vector into the channel, causing decoding errors at the receiver. White-box: full knowledge of the decoder architecture and parameters and synchronization with the transmitter; black-box (limited-knowledge, transfer-based): only the (η,k) configuration and channel model are known, the target model cannot be accessed or queried, and transfer relies on surrogate models. The victim system is an end-to-end autoencoder communication system.
- End-to-end autoencoder communication systems are easily deceived by physical-layer adversarial perturbations: an attacker injecting a low-power perturbation into the channel can cause decoding errors at the receiver, while existing physical attacks are mostly gradient methods, and effective attacks in black-box, knowledge-limited scenarios are understudied.
- The attack is modeled as an optimization problem of maximizing BLER under a power constraint, solved with an evolutionary algorithm: for white-box, PAGA (genetic algorithm + elite initialization) is used; for black-box, PAMFGA (multi-factorial evolution jointly optimizing over multiple surrogate autoencoders, then aggregating a consensus perturbation via SVD to transfer to an unknown target).
- Simulations show that PAGA has higher BLER than FGMA, RMAEP and Gaussian jamming under all (η,k) and PSR settings; under black-box conditions, PAMFGA outperforms running PAGA individually and then aggregating in both BLER and runtime; the perturbations transfer across architectures (including CNN) and remain effective under simulated conditions such as fading and CSI error.
- The authors state that the study remains at the simulation level, with no hardware over-the-air validation, and covers only a limited set of communication configurations, architectures and surrogate models; under black-box conditions, PAMFGA's advantage over baselines is small at some operating points; at T=1, PAMFGA has no advantage.
- Evaluated models
- MLP-based autoencoderMLP_reluMLP_deepMLP_widedecCNN-based autoencoder
- Benchmarks
- AWGN 信道下 (η,k) ∈ {(5,3),(7,4),(11,6)} 的仿真自编码器通信系统Rayleigh/Rician 衰落、CSI 误差、同步失配、CFO、定时偏移、路径损耗等仿真信道条件
- Metrics
- BLERPSR运行时间
The evolutionary-algorithm-based physical attack PAGA and its multi-factor version PAMFGA can significantly reduce the reliability of end-to-end autoencoder communication systems in the evaluated settings. PAGA targets the white-box setting, combining a genetic algorithm with elite population initialization to construct attack signals that induce significant reconstruction error and misclassification; PAMFGA targets the black-box setting where the attacker has limited knowledge of the receiver. The study emphasizes that the design and implementation of such systems must incorporate security considerations.
Paper analysis
What problem does this paper attempt to solve?
How to launch power-constrained physical-layer adversarial attacks against end-to-end autoencoder communication systems (white-box and black-box settings).
Core question: how to construct, under a power constraint, a universal physical adversarial perturbation that can propagate through a wireless channel and cause decoding errors in an end-to-end autoencoder communication system.
- Problem setting and importance: the authors consider DL autoencoder communication a 6G candidate technology, but small perturbations to the channel input propagate through encoder—channel—decoder and raise the receiver's BLER; compared with conventional jamming, adversarial perturbations have low energy and small spectral signatures.
- Shortcomings of existing methods: the authors state that existing physical attacks (e.g., FGMA, RMAEP) mainly rely on gradient optimization, and that the literature mostly focuses on digital attacks under idealized assumptions; knowledge-limited black-box physical attacks "remain relatively understudied."
- Core observation/hypothesis: the BLER objective is non-convex, non-smooth, and stochastic, and gradients are unstable or unavailable, making derivative-free evolutionary algorithms suitable; the black-box case relies on the transferability of adversarial examples—a perturbation constructed for one model remains effective for other models on the same type of task.
- What the paper proposes: the white-box attack PAGA (genetic algorithm + elite initialization) and the black-box attack PAMFGA (multi-factorial evolution + SVD aggregation transfer), along with an optimization formulation for maximizing BLER.
- Threat model:
- Objective: inject a perturbation vector p so that the decoded output ŝ≠s, i.e., maximize BLER, with perturbation energy constrained by PSR.
- White-box knowledge: full knowledge of the decoder DNN's architecture, weights, and biases, and synchronization with the transmitter; the authors state this is a controlled full-knowledge benchmark setting.
- Black-box knowledge: only the system-level configuration (η,k) and the physical channel model are known; the target architecture/parameters cannot be accessed and the target model cannot be queried (limited-knowledge, transfer-based).
- Victim system: point-to-point end-to-end autoencoder communication systems (MLP/CNN autoencoders, AWGN and fading channels).
What related research exists?
Related work falls into three groups: autoencoder communications, wireless adversarial attacks, and evolutionary/multi-factorial optimization; the baselines are FGMA, RSBA, RMAEP, and Gaussian jamming.
End-to-end autoencoder communication
- O'Shea & Hoydis [29]—propose end-to-end DL autoencoder physical-layer communication with joint transmitter and receiver optimization, which is the basis of this paper's system model.
- SCMA direction [30, 31, 32]—extend the autoencoder paradigm to code-domain NOMA: learning codebooks and decoding, reducing decoding complexity, and combining with Wasserstein GAN to improve noise robustness.
- Component-level DL applications [24, 25, 26]—DL for resource allocation, channel estimation, power control, signal detection, etc.
Adversarial attacks in wireless systems
- Sadeghi & Larsson [34, 42]—[34] adversarial attacks against DL wireless signal classification; [42] proposes the physical attacks FGMA (white-box gradient method) and RSBA (black-box surrogate-model method) against autoencoder communication, which are the direct comparison targets of this paper.
- Son et al. (RMAEP) [38]—a projected-gradient physical attack against dual-RIS-assisted MIMO autoencoders; in this paper's black-box experiments it is adapted into an SVD aggregation variant for comparison.
- Papernot et al. [39]—research on the transferability of adversarial examples, which is the theoretical basis for this paper's black-box setting.
- The paper notes: the literature mostly emphasizes digital attacks under idealized assumptions [34], which may not transfer to real RF environments [38]; knowledge-limited black-box physical attacks are understudied.
Evolutionary and multi-factorial optimization
- EA/GA [40, 41]—derivative-free, population-based stochastic optimization, suitable for non-convex, multimodal problems; the authors consider GA's diversity and crossover/mutation operators suitable for constructing physical perturbations, superior to the recombination flexibility of PSO [48] and DE [49].
- MFEA / MFEA-II [50, 58]—multi-factorial evolution places multiple related tasks in a unified population for implicit knowledge transfer; MFEA-II's online RMP estimation is the basis of PAMFGA.
Baselines and positioning
- Baseline methods: FGMA, RSBA, RMAEP (SVD-adapted variant in the black-box case), Gaussian jamming; additionally GA [59], DE [60], PSO [61], CMA-ES [62] are compared under a matched evaluation budget.
- Positioning: the authors state that the novelty of this paper is not in GA or transfer learning itself, but in specializing evolutionary optimization to power-constrained universal physical perturbation design (PAGA), and in the two-level transfer mechanism of inter-task transfer across multiple surrogate models plus cross-model transfer via the SVD consensus direction (PAMFGA) (Table 1).
How does the paper address this problem?
The attack is formalized as maximizing BLER under a power constraint, solved using a genetic algorithm (PAGA) and multi-factor evolution + SVD aggregation (PAMFGA).
Overall approach: the physical attack is written as a constrained optimization problem, and a derivative-free evolutionary algorithm is used to search for the perturbation. PAGA handles the white-box case, and PAMFGA uses multi-factorial evolution and two-level transfer to handle the black-box case.
Problem formulation
- System: message s∈M (|M|=2^k) is mapped by the encoder to an η-dimensional complex signal x, passes through an AWGN channel, and the receiver's decoder g outputs an estimate; the attacker injects p, and the received signal is y=x+n+p.
- Attack objective: maxp∈ℂη BLER(p), s.t. |p|22 ≤ 𝒫max, i.e., maximize the block error rate within the perturbation power budget; BLER is taken as an expectation over the message distribution and noise.
- The authors state that this objective is non-convex, non-smooth, and stochastic, with no closed-form solution or global optimality guarantee, so a population-based derivative-free evolutionary algorithm is used.
PAGA (white-box)
- Solution representation: the chromosome is a 2η-dimensional real vector (real/imaginary parts of the complex perturbation); when out of bounds, it is normalized and projected back onto the power-constraint sphere per Eq. (16).
- Initialization: a mixture of three types—uniform random, normal random (simulating CN(0,Pmax)), and elite initialization (for m randomly selected symbols, a fast gradient method is used to generate perturbations as "elite" seeds).
- Evolutionary operators: crossover uses HX (three-segment: both ends inherited, middle segment averaged) and SBX (distribution index ηc=2); mutation uses CIM (reversing gene order around a random pivot) and PM (polynomial mutation); re-normalization is performed after each mutation.
- Fitness and selection: for all M equiprobable messages, one noise realization is sampled each, and the proportion of decoding errors is counted as the approximate BLER (Algorithm 1); individuals are ranked by fitness, combined with age-based elimination and elitism (elite proportion pe), and the best individual is output after a fixed Gmax generations.
PAMFGA (black-box)
- Surrogate models: the attacker uses the known (η,k) and channel model to independently train T MLP surrogate autoencoders offline; each surrogate model corresponds to one attack task, sharing the same perturbation search space (USS).
- Multi-factorial evolution: based on MFEA-II, individuals carry a skill factor; the RMP matrix controls the cross-task mating probability (initial off-diagonal entries 0.7, with 0.95 given as another example in the main text), adaptively increased or decreased depending on whether offspring outperform parents (Eq. 32); crossover uses HX/SBX plus dOX (window length modulated by RMP), mutation is the same as PAGA, and selection is by scalar fitness.
- SVD aggregation transfer: after evolution ends, the elite perturbations of the T tasks are taken, normalized, and assembled into a matrix for SVD; the first right singular vector v1=argmax|v|=1∑j |qjH v|2 is the consensus direction with the greatest alignment to the elite perturbations of all surrogate models, and is rescaled to the power constraint for attacking the unknown target.
- Two-level transfer (Remark 3): the first level is knowledge transfer between tasks via RMP during evolution; the second level is cross-model transfer to the inaccessible target after SVD aggregation. The authors state that SVD aggregation is a low-overhead, target-agnostic heuristic and does not guarantee being strongest against every unknown target.
Key parameters and complexity
- Hyperparameters: Popsize=100, Gmax=250, crossover probability 0.8, mutation probability 0.2, elite rate 5%, ηc=2 (after pre-tuning).
- Complexity: PAGA per generation is dominated by fitness evaluation O(Popsize·2^k·C_N); PAMFGA additionally incurs training of T surrogate models O(T·C_train) and SVD O(ηT·min{2η,T}).
What experiments did the paper conduct?
Simulation evaluation: white-box PAGA achieves the highest BLER across all settings; black-box PAMFGA outperforms running PAGA individually followed by SVD aggregation in both BLER and runtime when T≥4.
Experimental setup
- Systems under test: MLP autoencoders, (η,k)∈{(5,3),(7,4),(11,6)}; cross-architecture experiments use MLP_base as the source and MLP_relu/MLP_deep/MLP_widedec/CNN as targets; the main black-box experiment target is a CNN autoencoder.
- Channels: AWGN by default; robustness experiments add Rayleigh, Rician (different K factors), CSI error, synchronization mismatch, CFO, timing offset, path loss, and a combined OTA-inspired simulation.
- Attack parameters: PSR∈{−2,−6,−10} dB; SNR=1 dB during training and fitness evaluation, with performance curves sweeping Eb/N0.
- Baselines: FGMA, RSBA, RMAEP (SVD-adapted variant in the black-box case), Gaussian jamming; additionally GA, DE, PSO, CMA-ES are compared (with matched fitness evaluation budget).
- Metrics and protocol: BLER (proportion of decoding errors); sensitivity experiments use 10 independent runs with 95% CI, and fresh BLER uses 1000 new noise realizations per message (16,000 samples in total).
- Training: each autoencoder is trained for 1000 epochs, batch 1000, Adam with initial learning rate 3×10⁻⁴, cosine decay.
Main results
Main black-box results (Table 3, target is a CNN autoencoder, BLER versus Eb/N0, excerpt):
Swipe horizontally to read the table
Method Eb/N0=0 4 8 12 15 PAMFGA (T=14) 0.3364 0.1245 0.0452 0.0197 0.0119 RSBA (T=14) 0.3072 0.1242 0.0445 0.0148 0.0058 PAGA (T=14, SVD aggregation) 0.3332 0.1245 0.0430 0.0152 0.0064 PAMFGA (T=4) 0.2876 0.0963 0.0216 0.0039 0.0004 - The authors state that in the white-box case, PAGA achieves the highest BLER across all (η,k) and PSR settings in Figs. 4a–4i: at PSR=−2 dB, (7,4), PAGA is about 2×10⁻¹ while jamming is about 9×10⁻³; at PSR=−6 dB, (5,3), PAGA is about 10⁻³ while jamming is about 10⁻⁵.
- In the black-box case, PAMFGA has the highest or near-highest BLER at most Eb/N0 points, but the authors note that at some operating points the gap to RSBA and the SVD-adapted PAGA/RMAEP is small, and that the BLER axis is on a logarithmic scale, so visual gaps do not represent large absolute gaps.
- Table 4 (Eb/N0=14): at T=1, PAMFGA (6e-5) is slightly lower than PAGA (7e-5) and slower (1.94 vs 1.14 minutes); at T=4, both have the same BLER (5e-4) but PAMFGA is faster; at T=6/10/14, PAMFGA has higher BLER and is faster (e.g., T=14: 0.0140 vs 0.0087, 11.34 vs 18.78 minutes).
Cross-architecture transfer
- What was tested: perturbations constructed on MLP_base are transferred directly to 4 unseen architectures (Eb/N0=4 dB, PSR=−6 dB, Fig. 6).
- Results: BLER for MLP variants rises from about 0.025 to 0.074–0.080; for the CNN target it rises from about 0.024 to 0.087.
- Authors' interpretation: the perturbations capture features that are effective across architectures, empirically supporting the black-box transfer assumption; the authors note that the degradation for the out-of-family CNN is comparable to or even larger than for in-family targets.
Robustness and convergence
- Robustness (Figs. 9, 10): under Rayleigh/Rician fading, CSI error, synchronization mismatch, CFO, timing offset, and path loss, the marginal BLER increase of the transfer attack remains positive, and PAMFGA is generally the largest among black-box methods; however, under strong path loss the marginal increase of all methods becomes smaller. The authors state that these are still simulations and do not represent hardware over-the-air validation.
- Convergence (Fig. 8): PAGA fitness rises rapidly for about the first 50 generations and levels off after about 100 generations; PAMFGA rises rapidly for the first 20–40 generations and enters a plateau after about 80–100 generations.
Other ablations and analyses
- Random fitness sensitivity (Table 2, (7,4), PSR=−6dB): fresh BLER for Nfit=1/5/10 is 0.2417/0.2407/0.2458, with paired-difference 95% CIs all containing 0; the selection optimism bias drops from 0.4396 to 0.1417, and runtime rises from 11.14 s to 68.55 s; on this basis the authors keep Nfit=1.
- Matched-budget comparison (Fig. 5a, PSR=−6dB): PAGA has higher BLER than GA, DE, PSO, and CMA-ES at most Eb/N0 points; Fig. 5b shows PAGA's runtime is comparable to the other metaheuristics, while FGMA/RMAEP are faster; the authors state they do not claim PAGA is the fastest.
- Number of surrogate models T (Table 3, Remark 4): increasing T from 4 to 10 raises BLER at most SNR points, and the gap between T=10 and T=14 becomes clearly smaller; the authors consider the benefit to saturate after about T=10, but do not claim T=10 is globally optimal.
What are some points that can be further explored?
The authors state that the study is limited to simulation with limited configurations; future work includes over-the-air hardware validation, broader scenarios and defense research.
Limitations and future directions noted by the authors
- Simulation level: The study is still simulation-based and considers only a limited set of communication configurations, architectures and surrogate models (Section 7).
- Hardware validation: Hardware over-the-air validation is left for future work (Section 7).
- Broader scenarios: Broader MIMO/RIS-assisted and mobile scenarios, and more diverse target models, are left for future work (Section 7).
- Defense directions: Future work will study adversarial training, perturbation detection, robust autoencoder design and joint attack-defense optimization (Section 7).
- Idealized white-box assumptions: Perfect synchronization and perfect channel knowledge may not hold in actual over-the-air deployments; timing/carrier frequency offset and channel estimation error may reduce attack effectiveness, and white-box results should be interpreted as a controlled full-knowledge benchmark (Remark 1).
- No guarantee for SVD aggregation: SVD aggregation is a target-agnostic heuristic; it does not guarantee the strongest attack against every unknown target, nor does it claim to outperform all aggregation strategies (Section 5.3).
Experimental coverage
- The systems under test are MLP autoencoders (three (η,k) configurations) and a CNN autoencoder, with 4 cross-architecture targets in total (MLP_relu, MLP_deep, MLP_widedec, CNN).
- Channels cover AWGN, Rayleigh, Rician, CSI error, synchronization mismatch, CFO, timing offset, path loss and combined OTA heuristic simulation, all simulated.
- Sensitivity experiments use 10 independent runs per setting and report 95% CI; the paper does not uniformly report the number of repetitions for other experiments.
- The surrogate model for black-box experiments is an independently trained MLP, T∈{1,4,6,10,14}; the target architecture differs from the surrogate architecture (CNN), but the surrogate models differ only in initialization and training data.
- The paper does not report the total number of queries/evaluations required for attack construction (only the fitness evaluation with M simulations per individual per generation and the runtime).
Summarize the main content of the paper
Physical-layer adversarial attacks against autoencoder communication: PAGA (white-box) and PAMFGA (black-box multi-surrogate transfer), with BLER in simulation higher than existing attacks.
A paper on physical-layer adversarial attacks against end-to-end autoencoder communication systems, modeling the attack as an optimization problem of maximizing BLER under a power constraint and solving it with an evolutionary algorithm.
- Problem: DL autoencoder communication is susceptible to low-power physical perturbations; existing physical attacks are mostly gradient methods, and knowledge-limited black-box scenarios are understudied.
- Method: white-box uses PAGA—a genetic algorithm with uniform/normal/fast-gradient elite mixed initialization, HX+SBX crossover, CIM+PM mutation, evolving a universal perturbation by fitness (approximate BLER); black-box uses PAMFGA—treating T independently trained surrogate autoencoders as multiple tasks for MFEA-II-style multi-factorial evolution (RMP-adaptive cross-task mating), then using SVD to extract a consensus direction from the elite perturbations of each task and transfer it to the unknown target, forming a two-level transfer of inter-task plus cross-model.
- White-box results: across all (η,k)∈{(5,3),(7,4),(11,6)} and PSR∈{−2,−6,−10} dB settings, PAGA's BLER is higher than FGMA, RMAEP, and Gaussian jamming (e.g., at PSR=−2 dB, (7,4), about 2×10⁻¹ vs jamming about 9×10⁻³); under a matched evaluation budget it is also higher than GA, DE, PSO, and CMA-ES, with runtime comparable to the other metaheuristics.
- Black-box results: against the CNN target, PAMFGA has the highest or near-highest BLER at most SNR points; at T=14, Eb/N0=14, BLER is 0.0140 with a runtime of 11.34 minutes, better than running PAGA individually followed by SVD aggregation (0.0087, 18.78 minutes); at T=1, PAMFGA has no advantage.
- Transfer and robustness: perturbations constructed on MLP_base transferred to an unseen CNN target raise BLER from about 0.024 to 0.087; the attack remains effective under simulated conditions including Rayleigh/Rician fading, CSI error, synchronization mismatch, CFO, and path loss, but under strong path loss the marginal effect of all methods becomes smaller.
- Authors' conclusions: high nominal reliability does not imply adversarial robustness, and security considerations should be incorporated into the design of autoencoder communication systems; the study is limited to simulation, with hardware over-the-air validation, broader MIMO/RIS scenarios, and defenses such as adversarial training and perturbation detection left for future work.