The Fast Gradient Sign Method (FGSM) has become such a critical threat of adversarial attacks on deep learning models for processing medical images. The problem has led to prediction errors with non-satisfactory results on diagnosis and patient safety. We address this challenge by proposing a novel approach named Genetic Algorithm-based Adaptive Compression (GA-AC) for recovering images perturbed by the FGSM attacks. The GA-AC optimize PNG and WebP compression methods to maximize Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM), thereby preserving essential diagnostic features in the restored images. Experimental results on multiple X-ray images demonstrate the effectiveness of GA-AC, which was able to restore the model’s F1-score from 24.14% to 98.10% after FGSM attacks.
Lima et al. (Mon,) studied this question.