Evaluation of RPL Preinstalled Secure Mode Under Attacks

Deep learning has revolutionized medical imaging by enabling automated disease diagnosis with high accuracy. But these models remain vulnerable to adversarial attacks, where imperceptible perturbations to input images can lead to incorrect predictions. In this study, we evaluate the robustness of DL models against both direct and transfer adversarial attacks in medical imaging. Using the Oesophageal Adenocarcinomas dataset, we implement the FGSM to quantify the impact of adversarial perturbations on model performance. We propose and compare three defense mechanisms: Total Variation Minimization (TVM), Adversarial Training (AT), and a Binary Input Detector (BID). Our results demonstrate that BID reduces the attack success rate to 0%, while AT and TVM yield average attack success rates of 35% and 75%, respectively. However, evaluation under transfer attacks reveals a limitation: BID trained on adversarial examples with high epsilon struggles to filter low-perturbation attacks, resulting in a high attack success rate of up to 100%. Nonetheless, it remains effective in detecting medium and high perturbation attacks. In contrast, TVM and AT exhibit average attack success rates around 64% and 42% across all epsilon values, indicating they are less effective than BID.