Adversarial Robustness of Deep Learning in Medical Imaging: A Comprehensive Survey and Benchmark of State-of-the-Art Architectures

The integration of artificial intelligence into medical diagnostics promises to revolutionize healthcare. However, the reliability of these systems is critically undermined by adversarial examples, which are imperceptible perturbations that can lead to misdiagnosis. Ensuring the robustness of AI-driven clinical decisions is paramount for ensuring patient safety and institutional trust. This study addresses this challenge in two ways. First, we provide a structured survey of the state-of-the-art adversarial threats, including adversarial attacks and detection strategies. Second, we present a rigorous empirical benchmark of five prominent CNN architectures for dermatoscopic skin cancer classification using the gold standard Auto Attack suite. The results revealed significant disparities in robustness based on the architectural design. Although all standard-trained models are highly vulnerable, their defensibility through adversarial training varies significantly. We found that modern transformer-inspired architectures, such as ConvNeXt, achieved the state-of-the-art robust accuracy while maintaining high performance with minimal trade-offs. Conversely, architectures optimized for mobile efficiency, such as MobileNetV2 and EfficientNet-B2, are exceptionally difficult to defend. To the best of our knowledge, this is the first study to establish an architectural hierarchy of robustness for dermatoscopic tasks, demonstrating that hybrid designs outperform mobile-optimized models by over 25% under adversarial conditions. These findings advocate a shift in clinical AI validation from accuracy-centric to robustness-centric metrics.