Towards building robust models for unimodal and multimodal medical imaging data

Joy Dhar, Puneet Goyal, Maryam Haghighat, Nayyar Zaidi, Ferdous Sohel, Quoc Bao Vo, KC Santosh

Abstract: Highlights•MI-Protector enhances adversarial robustness via example, model, and dual-level defenses.•DGMP integrates TLGAN and purifier for adversarial defense by generating purified images.•UMAN adds non-learnable and learnable UMANs at input and feature layers for defense.•MMIFAN uses model-level defense via attention noise injection at feature layers.•Experiments on 14 medical datasets verify our approach against adversarial attacks.

External IDs:doi:10.1016/j.inffus.2025.103822