ARB4WM: An Adversarial Robustness Benchmark for World Models in Continuous Control
Researchers have introduced ARB4WM, a new benchmark designed to evaluate the adversarial robustness of world models in continuous control systems. This framework assesses threats across policy, value, and latent dynamics levels, utilizing visual perturbations. The study demonstrates that attacks targeting value estimation and latent representations can be as detrimental as direct policy disruptions, highlighting the need for comprehensive safety assessments that consider multiple attack objectives and temporal exposure protocols. AI
IMPACT Introduces a new benchmark for evaluating the safety and robustness of AI world models in continuous control systems.