PhD Forum: Learning Adaptive Adversarial Attacks for Medical AI via Reinforcement Learning

Adversarial attacks on medical image classifiers typically use fixed hyperparameters, applying the same perturbation budget uniformly regardless of image difficulty. This is suboptimal: an intelligent attacker seeks the minimum perturbation sufficient for misclassification while minimizing queries to avoid detection. We present MedSecure, a hierarchical reinforcement learning framework that learns adaptive, per-image attack policies. A Double DQN meta-controller selects the attack strategy while independent Soft Actor-Critic controllers optimize continuous parameters conditioned on each image. A multi-objective reward function, gating success on both misclassification and perceptual quality, drives the agent to minimize perturbation magnitude rather than defaulting to a fixed budget. The domain-agnostic state representation, capturing gradient statistics, model confidence, attack progress, and image characteristics, is designed to enable zero-shot transfer across medical imaging domains and victim architectures.