Woojae presented “RoME: Robust Mixture of Low-Rank Experts against Multiple Adversarial Perturbations” at ECCV 2026.
RoME introduces a mixture-of-experts framework for multi-perturbation adversarial training, using low-rank expert updates to separate threat-specific features from shared representations. It further uses dual-scale gating and threat-guided routing diversification to induce distinct expert pathways across perturbation types, improving union robustness, natural accuracy, and robustness to unseen threats.