Optimization of Communication Signal Adversarial Examples by Selectively Preserving Low-Frequency Components of Perturbations

Yi Zhang, Lulu Wang, Xiaolei Wang, Dianxi Shi, Jiajun BaiSensors 2024 · DOI

TL;DR

This paper introduces a novel method to create more stealthy adversarial examples for communication signal recognition systems. By focusing on preserving only the low-frequency components of the perturbations, they significantly reduce distortion while maintaining or even slightly improving the attack's success rate. This approach offers a more efficient way to generate effective adversarial attacks that are harder for humans to detect.

What Problem Does This Paper Address?

Adversarial examples are subtly modified inputs designed to fool machine learning models, particularly deep neural networks (DNNs). In the realm of communication signal recognition, generating these adversarial examples is crucial for understanding the robustness and security of these systems. The core challenge lies in striking a balance between two competing goals: achieving a high Attack Success Rate (ASR), meaning the model misclassifies the signal, and minimizing perturbed distortion, making the adversarial example imperceptible or indistinguishable from the original signal. Existing methods often struggle to achieve both simultaneously, leading to either highly noticeable perturbations or attacks that are too weak to be effective. This paper aims to bridge this gap by optimizing adversarial examples for communication signals to be both highly effective and minimally distorted.

Methods

The paper's core innovation lies in its focus on the Discrete Cosine Transform (DCT) domain and the selective preservation of Low-Frequency Components of Perturbations (LFCP). The authors observed that DNNs used for communication signal recognition pay significant attention to the low-frequency components of the signal's DCT representation. They hypothesize that by carefully manipulating these low-frequency components of the perturbation, they can effectively alter the model's predictions while keeping the overall distortion low.

Their proposed method involves:

  1. DCT Transformation: Converting the original signal and its adversarial counterpart into the DCT domain.
  2. Identifying LFCP: Using a binary search algorithm to find an optimal "index threshold." This threshold determines which DCT coefficients of the perturbation are considered "low-frequency" and should be preserved.
  3. Selective Preservation: Applying a mask to the perturbation's DCT coefficients, keeping only the identified LFCP and discarding the high-frequency components.
  4. Inverse DCT: Transforming the modified perturbation back into the time domain and adding it to the original signal to create the optimized adversarial example.

The binary search algorithm is crucial here, as it iteratively searches for the smallest index threshold that ensures the optimized example is misclassified by the DNN, while simultaneously minimizing the overall perturbation.

Datasets

  • RadioML2018.01A: This is a publicly available dataset containing IQ signals for 24 different modulation types across 26 signal-to-noise ratios (SNRs). It was used for training and evaluating the DNN models. The dataset was partitioned into training and testing sets with a 7:3 split.

Key Insights and Ideas

  • DNNs are sensitive to low-frequency perturbations: The research strongly suggests that DNNs used for communication signal recognition are particularly influenced by the low-frequency components of adversarial perturbations. This is a key insight that drives the proposed optimization strategy.
  • DCT domain is informative: Analyzing perturbations in the DCT domain reveals their distribution and highlights the importance of low-frequency coefficients for altering model predictions.
  • Selective preservation is key: Instead of adding arbitrary perturbations, the method's effectiveness stems from selectively keeping only the low-frequency components of the perturbation. This allows for significant distortion reduction without sacrificing attack efficacy.
  • Binary search for optimal trade-off: The use of a binary search algorithm is an elegant way to find the best balance between minimizing distortion (low RMSE) and ensuring misclassification (high ASR) by efficiently searching for the optimal LFCP threshold.

Key Results

  • Significant reduction in perturbed distortion: The proposed method consistently minimized the Root Mean Square Error (RMSE) of adversarial examples. For FGSM, BIM, PGD, and MI-FGSM, the average minimization in ARMSE (Average RMSE) was substantial, ranging from 55.6% to 61.9% across different models and perturbation parameters.
  • Modest improvement in ASR: While the primary goal was distortion reduction, the optimized adversarial examples often achieved a slight improvement in ASR compared to the original adversarial examples. This improvement was more pronounced for attacks like FGSM, BIM, and PGD.
  • Feasible performance on challenging attacks: Even for attacks like DeepFool and BS-FGM, which are designed to produce very small perturbations and already achieve high ASRs, the proposed method still demonstrated a positive impact, further reducing distortion.

Limitations and Future Work

  • No direct comparison to existing optimization methods: The authors note that comparative experiments with other optimization methods for communication signal adversarial examples were not conducted due to the absence of such existing methods in the literature.
  • Generalizability to other signal types: While the paper suggests the method is not limited to a specific signal style and could be applied to images and audio, this remains a future exploration.
  • Exploring other transformations: The authors propose investigating the use of other time-frequency transformation approaches like Discrete Fourier Transform (DFT) and Discrete Wavelet Transform (DWT) for optimizing adversarial examples.
  • Parameter tuning: The effectiveness of the method is influenced by the choice of the index threshold n. While a binary search is used to find an optimal n, further research into automatic or adaptive methods for determining this threshold could be beneficial.

Summary updated

Abstract

Achieving high attack success rate (ASR) with minimal perturbed distortion has consistently been a prominent and challenging research topic in the field of adversarial examples. In this paper, a novel method to optimize communication signal adversarial examples is proposed by focusing on low-frequency components of perturbations (LFCP). Observations on model attention towards DCT coefficients reveal the crucial role of LFCP within adversarial examples in altering the model's predictions. As a result, selectively preserving LFCP is established as the fundamental concept of the optimization strategy. By utilizing the binary search algorithm, which considers the inconsistency in the model's predictions as a constraint, LFCP can be effectively identified, and the aim of minimizing perturbed distortion while maintaining ASR can be achieved. Experimental results conducted on a publicly available dataset, six adversarial attacks and two DNN models, indicate that the proposed method not only significantly minimizes perturbed distortion for FGSM, BIM, PGD, and MI-FGSM but also achieves a modest improvement in ASR. Notably, even for DeepFool and BS-FGM, which introduce small perturbations and exhibit high ASRs, the proposed method can still deliver feasible performance.
BibTeX
@inproceedings{zhang2024optimization,
  author = {Yi Zhang and Lulu Wang and Xiaolei Wang and Dianxi Shi and Jiajun Bai},
  title = {{Optimization of Communication Signal Adversarial Examples by Selectively Preserving Low-Frequency Components of Perturbations}},
  booktitle = {Sensors},
  year = {2024},
  doi = {10.3390/s24227110},
}
Figure