Adversarial Sample Generation Method for Modulated Signals Based on Edge-Linear Combination

Tianyi Wei, Yong Li, Yingtao NiuElectronics 2025 · DOI

TL;DR

This paper introduces a novel method for generating adversarial samples for wireless communication signals, specifically targeting modulation recognition systems. By using an "edge-linear combination" approach inspired by Mixup, they create subtle yet effective perturbations that can fool deep learning models, offering a more stealthy and efficient attack than traditional gradient-based methods.

What Problem Does This Paper Address?

Wireless communication systems are increasingly reliant on deep learning for tasks like modulation recognition (identifying the type of signal being transmitted, e.g., AM, FM, QAM). However, this reliance creates a vulnerability: sophisticated adversaries could use jamming techniques that exploit these deep learning models. If a jammer can accurately identify a system's modulation scheme, it can craft highly targeted interference, leading to communication blackouts. This paper addresses the challenge of creating "adversarial samples" – subtly modified signals designed to trick these modulation recognition models into misclassifying them. The goal is to disrupt the jammer's ability to perform targeted interference, thereby protecting normal communication. Existing methods often rely on direct access to model gradients, which isn't always feasible in real-world scenarios, and can sometimes produce less stealthy perturbations.

Methods

The core of the proposed method is the "edge-linear combination" technique, inspired by the Mixup data augmentation strategy. Instead of directly calculating gradients to find perturbations, this approach focuses on creating new samples by linearly combining existing ones.

Here's a breakdown:

  • Edge-Linear Combination: The method takes two original signal samples, $x_i$ and $x_j$, and creates a new sample $x_{adv}$ using a weighted sum: $x_{adv} = (1 - \alpha)x_i + \alpha x_j$. The key is that $\alpha$ is kept very small (between 0.05 and 0.07), ensuring $x_{adv}$ is very close to one of the original samples ($x_i$). This proximity to the "decision edge" of the neural network is where adversarial perturbations are hypothesized to be most effective.
  • Iterative Perturbation Selection: The process involves iterating through various combinations of original samples and their corresponding edge-linear combinations. The neural network is used to classify these generated samples. By counting how often a specific combination leads to a misclassification (an "error"), the method identifies the "best perturbation signal" for each original sample that is most susceptible to this type of attack.
  • Adversarial Sample Generation: Once the optimal perturbation signal is identified for a given original sample, the edge-linear combination is applied to generate the final adversarial sample. This process is designed to be efficient, especially for black-box scenarios where model gradients are inaccessible.

Datasets

The experiments utilize the RADIOML2016.10A dataset. This dataset contains 220,000 signal samples across 11 modulation modes (3 analog, 8 digital) and 20 signal-to-noise ratios (SNRs) ranging from -20 dB to 18 dB. Each signal sample includes in-phase (I) and quadrature (Q) data of length 128. The paper notes that for their experiments, they focus on high SNRs to ensure reliable results and preprocesses the data using min-max normalization to handle amplitude differences between modulation types.

Key Insights and Ideas

  • Decision Boundary Proximity is Key: The paper strongly suggests that adversarial samples are most effective when they lie very close to the decision boundaries of the neural network, even with minimal perturbations. The "edge-linear combination" method directly targets these regions.
  • Stealthy Attacks are Possible: By using small perturbation coefficients ($\alpha$), the generated adversarial samples are visually and statistically very similar to the original signals, making them harder to detect. This is a significant advantage over methods that might introduce more noticeable changes.
  • Gradient-Free Attack Viability: The proposed method demonstrates that effective adversarial attacks can be mounted without needing direct access to the neural network's internal parameters or gradients, making it applicable in more realistic black-box scenarios.
  • Targeted Perturbation Identification: The iterative process of identifying the "best perturbation signal" for each sample type allows for more precise and effective adversarial sample generation compared to generic noise addition or single-step gradient attacks.

Key Results

  • White-Box Attacks: The proposed edge-linear combination method achieved a higher attack success rate and a lower signal-to-noise ratio (SNR) compared to random noise adversarial samples with the same perturbation coefficient. It also reduced the neural network recognition accuracy by an average of 30% compared to traditional gradient-based methods under white-box attacks.
  • Black-Box Attacks: In black-box scenarios, the method showed a higher attack success rate and lower attack SNR compared to random noise adversarial samples.
  • Efficiency: While gradient-based methods like I-FGSM and MI-FGSM require significant computation time due to iterative gradient updates, the edge-linear combination method is faster once the optimal perturbation samples are identified, with a reported runtime of 0.08 hours compared to 0.81-0.9 hours for I-FGSM/MI-FGSM on the CLDNN model.
  • Robustness: The method's effectiveness was demonstrated on a different model (ResNet) than the one used for training (CLDNN), indicating good transferability of the generated adversarial samples.

Limitations and Future Work

  • Computational Cost for Optimal Perturbation: While generating adversarial samples is fast, the initial step of identifying the optimal perturbation signal can still be computationally intensive, especially with large datasets. The paper explores hyperparameter tuning (number of random samples) to balance this.
  • Attack Effectiveness vs. SNR: The paper notes that the edge-linear combination method's attack effect is optimal at higher SNRs (above -4 dB), suggesting it might be less potent in extremely noisy environments compared to gradient-based methods.
  • Future Work: The authors suggest that the generated adversarial samples can be used for adversarial training to improve the robustness of modulation recognition models. They also mention exploring other adversarial defense techniques like network distillation, adversarial detection, and network verification, indicating a broader research interest in both attack and defense strategies in this domain.

Summary updated

Abstract

In complex electromagnetic environments, wireless communication system reliability can be compromised by various types of jamming. To address the issue of jammers using deep neural network models to identify communication signal modulation method and apply targeted interference, this paper proposes a method for generating adversarial samples of modulation signals based on the Mixup linear combination approach. The method generates edge-linear combination samples with small perturbations by linearly combining the original signal samples near the decision edges, and then inputs them into the neural network model for identification test, determines the best perturbation signals for each type of signals according to the identification results, and then generates the adversarial samples by selecting the best perturbation signals for each type of modulation during the attack. Simulation results show that, compared to traditional gradient-based adversarial sample generation algorithms, the proposed method performs better under white-box attacks. Under black-box attacks, the proposed method achieves higher attack success rates and lower attack signal-to-noise ratios compared to random noise adversarial samples with the same disturbance coefficient.
BibTeX
@inproceedings{wei2025adversarial,
  author = {Tianyi Wei and Yong Li and Yingtao Niu},
  title = {{Adversarial Sample Generation Method for Modulated Signals Based on Edge-Linear Combination}},
  booktitle = {Electronics},
  year = {2025},
  doi = {10.3390/electronics14071260},
}
Figure