Abstract
Deep neural networks are highly effective for classification but remain vulnerable to subtle, adversarial noises. This paper proposes a novel defense mechanism that applies a recovery process to inputs before they reach a naturally trained classifier. By intentionally corrupting images with significant noise and then reconstructing them using SARGAN—a Generative Adversarial Network variant originally for radar signal denoising—we can effectively strip away adversarial effects. Our findings demonstrate that this preprocessing pipeline significantly restores the accuracy of standard classifiers under attack.
Introduction: The Vulnerability of Deep Learning
Despite achieving superhuman performance in computer vision, deep neural networks (DNNs) are notoriously fragile. Adversarial attacks—imperceptible, mathematically optimized perturbations—can force a model to misclassify an image with high confidence. While adversarial training is the standard defense, it is computationally expensive and often reduces a model's performance on clean data. This research explores an alternative: can we "wash" the data before it reaches the model?
Methods: The "Wash and Rinse" Defense
The authors propose a modular preprocessing pipeline that decouples defense from classification. This approach involves three primary stages:
- Intentional Corruption: Upon receiving an input, the system adds random white Gaussian noise. This step is counter-intuitive but crucial; it aims to "swamp" the precise adversarial perturbations with stochastic noise.
- Denoising with SARGAN: The corrupted image is processed by SARGAN (Synthetic Aperture Radar Generative Adversarial Network). Trained purely on standard noise patterns (not adversarial ones), SARGAN reconstructs the clean image.
- Sequential Filtering: To enhance the cleaning effect, the output can be passed through multiple SARGAN stages (up to four) sequentially.
Experimental Results
The defense was tested against Projected Gradient Descent (PGD) attacks on several benchmarks, including MNIST, Fashion-MNIST, and CIFAR-10/100. Without defense, PGD attacks reduced classifier accuracy to near 0%.
Applying a 4-stage SARGAN filter led to dramatic recoveries:
- MNIST: Accuracy rose from 0.66% to 89.08%.
- Fashion-MNIST: Accuracy rose from 0% to 71.76%.
- CIFAR-10: Accuracy rose from 0.99% to 51.58%.
Crucially, this preprocessing had minimal impact on clean data performance, typically maintaining accuracy within 1% of the baseline. This suggests that the "wash and rinse" cycle effectively targets malicious perturbations while preserving essential features.
Key Insights
- Fighting Noise with Noise: Stochastic Gaussian noise dilutes the specific, brittle structure of adversarial attacks.
- Attack Agnosticism: Because the denoiser is not trained on specific attacks, it remains a generic defense against various $L_p$ norm perturbations.
- Modularity: This method allows for the use of high-performance, naturally trained models without the overhead of adversarial retraining.
Conclusion
While the absolute performance of SARGAN-based recovery does not yet match the state-of-the-art in adversarial training, it offers a promising, modular alternative. Future work will focus on optimizing SARGAN specifically for adversarial filtering and testing against a wider array of attack vectors.
Summary updated