Self-Supervised Medical Adversarial Robustness With Two-Stage Contrastive Constraints

The burgeoning field of deep learning-based medical image analysis has witnessed remarkable advancements in revolutionizing clinical practice. Due to the inherent vulnerability of neural networks, prior studies inevitably suffer from a significant challenge of catastrophic consequences posed by adversarial attacks and perturbations, especially in critical applications such as clinical medicine. To address this issue, this paper proposes a novel Self-Supervised Medical Adversarial Robustness (SMAR) framework that jointly incorporates self-supervised contrastive learning with adversarial training to facilitate the robustness and generalization of medical diagnostic models. Specifically, the proposed SMAR framework is built upon a two-phase learning scheme, consisting of the contrastive pre-training (CPT) and subsequent adversarial fine-tuning (AFT) stages. During the CPT stage, self-supervised contrastive learning empowers the model to adaptively capture intrinsic structures and patterns from unlabeled data, thereby developing a powerful feature encoder capable of furnishing robust and informative representations for subsequent adversarial training. By contrast, the AFT stage mainly focuses on optimizing the adversarial training process by encouraging the model to adapt to both the natural variations present within the unlabeled data and the potential perturbations from adversarial samples. To the best of our knowledge, there is the first attempt to promote the adversarial robustness of medical diagnostic models, resulting in more reliable and robust diagnostic outputs. Extensive experiments on multiple medical benchmarks and natural scenes consistently demonstrate the superiority of our SMAR method over state-of-the-art baselines.