Adversarial Sample Generation Method Based on Frequency Domain Transformation and Channel Awareness

Yueming Gao, Dongwei Xu, Huiyan Zhu, Qi XuanSensors 2025 · DOI

TL;DR

This paper introduces SDRNet, a novel deep learning model for highly accurate channel estimation in OFDM systems, and a channel-aware adversarial attack method that leverages this improved estimation to significantly boost attack success rates in wireless communication security.

What Problem Does This Paper Address?

Accurate channel estimation is crucial for reliable wireless communication, especially in complex environments like those encountered in OFDM systems. Traditional methods struggle with low signal-to-noise ratios (SNRs) and channel sparsity. While deep learning has shown promise, existing methods can be limited. Furthermore, the security of these systems is threatened by adversarial attacks, which aim to trick machine learning models. Existing adversarial attacks often overlook the nuances of the wireless channel, limiting their effectiveness. This paper addresses both the need for more robust channel estimation and the development of more potent adversarial attacks that exploit channel characteristics.

Methods

The authors propose a two-pronged approach:

  1. SDRNet for Channel Estimation: They introduce the Super-Resolution Denoising Residual Network (SDRNet). This model combines the strengths of SRCNN (for super-resolution) and DnCNN (for denoising) with residual learning (inspired by ResNet) to enhance feature extraction and noise suppression. SDRNet is trained on OFDM pilot data with varying SNRs to learn robust channel characteristics. It's designed to handle frequency-selective fading channels effectively.
  2. Channel-Aware Adversarial Attack: Building on SDRNet's accurate channel estimation, they develop a frequency-domain adversarial attack method. This method first transforms the signal to the frequency domain using Fourier transforms. Then, it applies Gaussian noise and selective masking, crucially incorporating channel gradient information. This "channel awareness" allows the generated adversarial perturbations to be more targeted and effective, especially after traversing the actual wireless channel.

Datasets

  • Training Data: The paper mentions generating synthetic OFDM pilot data with varying signal-to-noise ratios (SNRs) to train the SDRNet model.
  • Evaluation Data: For adversarial attack evaluation, the authors utilize the publicly available RadioML2016.10A dataset. This dataset simulates complex wireless communication environments and includes 11 modulation types. It's a standard benchmark for modulation recognition tasks.

Key Insights and Ideas

  • Synergy of Super-Resolution and Denoising: Combining SRCNN and DnCNN architectures within SDRNet proves highly effective for channel estimation, addressing both low-resolution data and noise interference simultaneously.
  • Residual Learning for Robustness: Incorporating residual blocks (from ResNet) into the deep learning architecture significantly improves the model's ability to learn complex channel features and mitigates vanishing gradient issues, leading to better performance across various channel conditions.
  • Channel Awareness is Key for Attacks: Adversarial attacks that explicitly consider channel characteristics (like fading and noise) are far more effective than those that treat the channel as an ideal or static entity. This is because the channel itself can distort or amplify perturbations.
  • Frequency Domain is a Rich Attack Surface: Transforming signals to the frequency domain and applying targeted masks and noise allows for more sophisticated and potent adversarial perturbations.

Key Results

  • Channel Estimation Performance: SDRNet significantly outperforms traditional methods like LS and MMSE in terms of Mean Squared Error (MSE) and Bit Error Rate (BER). At 20 dB SNR, SDRNet achieved an MSE as low as 0.00004. Its performance remains robust across different SNRs and pilot interval schemes.
  • Adversarial Attack Performance: The proposed channel-aware adversarial attack method achieved a 79.9% attack success rate on the CNN1D model, which is 16.3% higher than non-channel-aware methods. This highlights the critical role of accurate channel estimation in crafting effective attacks. The method also demonstrated better perturbation concealment.

Limitations and Future Work

  • Generalization and Universality: The authors acknowledge that further research is needed to enhance the universality and concealment of adversarial samples, particularly under different channel conditions.
  • Complex Scenarios: Applying these methods to more complex communication scenarios and verifying their effectiveness and feasibility in real-world applications is suggested.
  • Data Balance: The authors note that the disturbance-to-signal ratio (PNR) in their adversarial attack experiments was not uniformly distributed, suggesting a need for dynamic sampling strategies in future work.

Summary updated

Abstract

In OFDM wireless communication systems, low-resolution channel characteristics and noise interference pose significant challenges to accurate channel estimation. To solve these problems, we propose a super-resolution denoising residual network (SDRNet), which combines the advantages of the super-resolution convolutional neural network (SRCNN) and the denoising convolutional neural network (DnCNN) to construct a pilot-based OFDM signal model, train SDRNet using OFDM pilot data containing Gaussian noise, and optimize its feature enhancement ability in frequency-selective fading channels. To further explore the role of channel estimation in communication security, we propose a frequency-domain adversarial attack method based on SDRNet output. This method first converts the time-domain signal to the frequency domain by using the Fourier transform and then applies Gaussian noise and selective masking. By integrating the channel gradient information, the adversarial perturbation we generated significantly improves the attack success rate compared with the non-channel awareness method. The experimental results show that SDRNet is superior to traditional algorithms (such as the least square method, minimum mean square error estimation, etc.) in both mean square error and bit error rate. Furthermore, the adversarial samples optimized through channel awareness frequency-domain masking exhibit stronger attack performance, confirming that accurate channel estimation can not only enhance communication reliability but also provide key guidance for adversarial perturbation. The experimental results show that under the same noise conditions, the MSE of SDRNet is significantly lower than that of LS and MMSE. The bit error rate is lower than 0.01 when the signal-to-noise ratio is 10 dB, which is significantly better than the traditional algorithm. The attack success rate of the proposed adversarial attack method reached 79.9%, which was 16.3% higher than that of the non-channel aware method, verifying the key role of accurate channel estimation in enhancing the effectiveness of the attack.
BibTeX
@inproceedings{gao2025adversarial,
  author = {Yueming Gao and Dongwei Xu and Huiyan Zhu and Qi Xuan},
  title = {{Adversarial Sample Generation Method Based on Frequency Domain Transformation and Channel Awareness}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25123779},
}
Figure