Reproduce “A Unified Framework for Adversarial Attack and Defense in Constrained Feature Space” on OpenResearch

About this paper

Thibault Simonetto, Salijona Dyrmishi, Salah Ghamizi, Maxime Cordy, Yves Le Traon

Researchers at the University of Luxembourg developed a unified framework to generate feasible adversarial examples and enhance model robustness in machine learning applications with complex domain constraints. Their multi-objective evolutionary attack, MoEvA2, achieved up to 100% success in creating constrained adversarial examples, and a novel defense using engineered non-convex constraints significantly improved model robustness, reducing MoEvA2's success rate to 0% in some cases.

Key insights

• Generic, constraint-unaware adversarial attacks are largely ineffective in real-world constrained domains, consistently failing to produce any feasible adversarial examples.
• Multi-objective evolutionary algorithms, such as MoEvA2, are highly effective for navigating complex, non-differentiable constrained search spaces to generate feasible adversarial examples.
• Augmenting machine learning models with engineered non-convex constraints provides a potent and complementary defense strategy against constrained adversarial attacks, sometimes surpassing or combining effectively with adversarial retraining.

Summary by alphaXiv.