Defense Against Adversarial Attacks with Efficient Frequency-Adaptive Compression and Reconstruction

The increasing use of deep neural networks exposes themselves to adversarial attacks in the real world drawn from closed-set and open-set, which poses great threats to their application in safety-critical systems. Since adversarial attacks tend to mislead an original model by adding small perturbations into clean images, an intuitive idea of defensing adversarial attacks is eliminating perturbations as much as possible to mitigate attacking effects. However, such elimination-based strategies unfortunately fail to achieve satisfactory robustness. Aiming to investigate the intrinsic reasons for this phenomenon, systematic experiments are carried out in this paper to indicate that even a 20% residual perturbation can still preserve and exhibit attacking effects as strong as a full one. Our study also indicates that there are strong correlations between perturbations and legitimate images. Thus, breaking the correlation across multiple bands is more effective in mitigating attacking effects. Based on these findings, this paper proposes an efficient defense strategy called "Frequency-Adaptive Compression and rEconstruction (FACE)" to improve the robustness of the model to adversarial attacks. Specifically, low-frequency bands containing semantic information are compressed by a down-sampling operation, while the channel width of high-frequency bands is squeezed and further compressed by adding noise before the Tanh activating function. Meanwhile, attachment spaces of perturbations are also squeezed to the extent as much as possible. Finally, a clean output is obtained by upsampling together with expanded reconstruction. Experiments are extensively conducted on widely used datasets to demonstrate the effectiveness of the proposed method. For closed-set attacks, FACE outperforms the STOA elimination-based methods on ImageNet, achieving a 27.9% improvement. For the MNIST open-set attacks, it not only reduces the success rate of targeted attack by a large margin (from 100% to 24.7%), but also mitigates attacking effects with an FPR-95 value of 0.3.