Population-Based Methods for Adversarially Resilient Deep Models

As deep learning systems become increasingly deployed in safety-critical applications, ensuring their robustness against adversarial threats has never been more vital. In this talk, I introduce a unified framework for adversarial resilience grounded in population-based methods - approaches that harness the diversity and structure of multiple models rather than relying on a single solution. We begin with mode connectivity techniques, including robust path and surface exploration, which construct low-loss connections between models trained under different threat models. These paths not only improve robustness across multiple Lp-norm perturbations but also enable efficient model ensembling and nonlinear machine unlearning. Building on this foundation, we explore how mixture-of-experts (MoE) architectures can incorporate population diversity through specialized expert training. We introduce a dual-model strategy and joint training mechanism that preserve accuracy while achieving certified adversarial robustness with minimal inference overhead. Together, these methods reveal how deep models can achieve robustness from the many via nonlinear parameter space structures, expert modularity, and flexible population dynamics.