TL;DR
Deep neural networks are notoriously vulnerable to adversarial attacks, where imperceptible noise causes catastrophic misclassification. This paper introduces a novel, counter-intuitive preprocessing defense: intentionally corrupting input images with random noise and then recovering them using a GAN-based denoiser (SARGAN) before classification. This "wash and rinse" cycle effectively dilutes the adversarial perturbations, significantly improving the robustness of naturally trained classifiers without requiring expensive adversarial retraining.
What Problem Does This Paper Address?
While deep learning models achieve superhuman performance on many vision tasks, they can be easily fooled by adversarial attacks—subtle, carefully crafted perturbations added to an image that are invisible to the human eye but force the network to misclassify it with high confidence.
The standard defense against this is adversarial training, where a network is trained on adversarial examples so it learns to ignore them. However, adversarial training is highly resource-intensive, time-consuming, and often degrades the model's accuracy on clean, uncorrupted images. The authors tackle this gap by asking: instead of changing the classifier to withstand attacks, can we build a generic preprocessing filter that "washes away" the malicious noise from the data before it ever reaches the classifier?
Methods
The authors propose a preprocessing pipeline that decouples adversarial defense from the classification model itself. The core approach involves three steps:
- Intentional Corruption: When an image (which may or may not contain an adversarial attack) is received, the system intentionally adds random white Gaussian noise to it.
- Denoising with SARGAN: The noisy image is passed through SARGAN (Synthetic Aperture Radar Generative Adversarial Network). Originally designed to denoise radar signals, SARGAN uses an encoder-decoder generator and a discriminator to reconstruct the original clean image. The network is trained purely on images corrupted with Gaussian noise or random mask cutouts, not on adversarial examples.
- Sequential Filtering: To maximize the cleaning effect, the authors chain multiple SARGAN denoisers together. The output of the first denoiser is fed into the second, and so on, up to four times.
- Classification: The reconstructed image is finally passed to a standard, naturally trained classifier.
To test the defense, the authors generate adversarial examples using the Projected Gradient Descent (PGD) attack, a powerful white-box method.
Datasets
The paper utilizes standard image classification benchmarks:
- MNIST (Handwritten digits)
- Fashion-MNIST (Clothing items)
- CIFAR-10 (10 classes of common objects)
- CIFAR-100 (100 classes of common objects)
No new datasets were introduced.
Key Insights and Ideas
- Fighting Noise with Noise: The idea of adding more noise to an already perturbed image seems counter-intuitive. However, the authors show that adding random Gaussian noise helps dilute or dominate the highly specific, mathematically precise adversarial perturbations. When the denoiser removes the Gaussian noise, it takes the adversarial noise out with it.
- Attack Agnosticism: Because the SARGAN denoiser is trained only to remove standard Gaussian noise and knows nothing about adversarial generation, it makes no assumptions about the specific type of attack (e.g., $L_p$ norm bounds). This makes it a highly generic defense mechanism.
- The Power of Sequential Denoising: A single pass through a denoiser helps, but chaining multiple denoisers sequentially acts as a much stronger filter, progressively flattening the adversarial effect with each pass.
- Preserving Clean Accuracy: By keeping the defense as a separate preprocessing step, researchers can use highly optimized, naturally trained classifiers without sacrificing much of their baseline performance on clean data.
Key Results
- Denoising Quality: SARGAN successfully reconstructed images corrupted by Gaussian noise, achieving average Peak Signal-to-Noise Ratios (PSNR) of 26.2 for MNIST and 26.4 for CIFAR-10.
- Adversarial Recovery: On naturally trained classifiers, PGD attacks successfully dropped accuracy to near 0% across all datasets. However, after passing the attacked images through a sequence of 4 SARGAN denoisers, accuracy rebounded dramatically:
- MNIST: Improved from 0.66% to 89.08%
- Fashion-MNIST: Improved from 0% to 71.76%
- CIFAR-10: Improved from 0.99% to 51.58%
- CIFAR-100: Improved from 0.69% to 29.29%
- Minimal Impact on Clean Data: Applying the 4-SARGAN preprocessing step to clean images only reduced the classifier's natural accuracy by a very small margin (mostly within 1 percentage point, except for a ~5% drop on Fashion-MNIST).
Limitations and Future Work
The authors honestly acknowledge that the absolute classification performance of this SARGAN recovery method is not yet on par with the absolute state-of-the-art robust models trained explicitly with adversarial training.
For future work, they plan to test this preprocessing method against other forms of adversarial attacks beyond PGD. They also suggest optimizing the SARGAN model specifically for adversarial filtering, which could close the performance gap and provide a highly competitive, modular alternative to adversarial training.
Summary updated