MoCo-EA: Exploiting Adversarial Mode Connectivity for Efficient Evolutionary Attacks

Yawen Zhao, Hyo Seo Kim, Gang Luo, Can Chen, Binghui Wang, Yue Duan, Ren Wang

Keywords: Adversarial attacks, evolutionary algorithms, mode connectivity, Bézier curves, transferability, genetic crossover

TL;DR: Continuous Bézier curves between adversarial perturbations outperform discrete crossover in evolutionary attacks, achieving universal success with drastically reduced computational cost.

Abstract: Evolutionary algorithms for adversarial attacks leverage population-based search to discover perturbations without gradient information, but suffer from inefficient crossover operations that destroy adversarial properties through discrete interpolation. We introduce Mode Connectivity Evolutionary Attack (MoCo-EA), which replaces traditional crossover with a novel Bézier crossover operator that optimizes perturbations along a continuous Bézier curve between parent perturbations. Our key insight is that adversarial examples lie on connected manifolds where intermediate points maintain, and often enhance, attack effectiveness. We demonstrate three findings: (1) Successful adversarial perturbations exhibit mode connectivity, forming continuous paths that preserve adversarial properties; (2) Intermediate points along optimized paths achieve higher transferability than endpoints, with improvements that scale with auxiliary image guidance; (3) Bézier crossover dramatically outperforms discrete genetic operations, achieving universal attack success across all perturbation norms while reducing convergence time and query requirements by orders of magnitude. By revealing the geometric structure of adversarial space and exploiting it through principled path optimization, MoCo-EA transforms evolutionary attacks from slow, unreliable processes into efficient, dependable methods. Our work challenges the traditional view of adversarial examples as isolated points and opens new directions for both attack generation and defense research.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 10744