Game-Theoretic Defenses for Robust Conformal Prediction Against Adversarial Attacks in Medical Imaging
Rui Luo, Jie Bao, Zhixin Zhou, Chuangyin Dang
TL;DR
The paper tackles adversarial threats in medical-imaging classifiers by integrating conformal prediction (CP) with a game-theoretic defense framework. It develops attack-aware defenses by training attack-specific models, calibrates CP thresholds under known and unknown perturbations, and exploits a zero-sum game to identify robust defensive strategies, including aggregation via Maximum and Minimum classifiers. The results on MedMNIST datasets show high CP coverage with relatively small prediction sets, with the game-theoretic approach often converging to a robust single defense and providing principled trade-offs between coverage and set size. This framework advances uncertainty quantification and adversarial robustness, offering a principled pathway for deploying reliable deep learning systems in adversarial healthcare environments.
Abstract
Adversarial attacks pose significant threats to the reliability and safety of deep learning models, especially in critical domains such as medical imaging. This paper introduces a novel framework that integrates conformal prediction with game-theoretic defensive strategies to enhance model robustness against both known and unknown adversarial perturbations. We address three primary research questions: constructing valid and efficient conformal prediction sets under known attacks (RQ1), ensuring coverage under unknown attacks through conservative thresholding (RQ2), and determining optimal defensive strategies within a zero-sum game framework (RQ3). Our methodology involves training specialized defensive models against specific attack types and employing maximum and minimum classifiers to aggregate defenses effectively. Extensive experiments conducted on the MedMNIST datasets, including PathMNIST, OrganAMNIST, and TissueMNIST, demonstrate that our approach maintains high coverage guarantees while minimizing prediction set sizes. The game-theoretic analysis reveals that the optimal defensive strategy often converges to a singular robust model, outperforming uniform and simple strategies across all evaluated datasets. This work advances the state-of-the-art in uncertainty quantification and adversarial robustness, providing a reliable mechanism for deploying deep learning models in adversarial environments.