AttackVLA: A Benchmark for Adversarial and Backdoor Attacks Against Vision-Language-Action Models

The researchers propose the AttackVLA framework, covering the full VLA development lifecycle of data construction, model training, and inference, uniformly evaluating adversarial and backdoor attacks against vision-language-action (VLA) models, implementing all existing VLA attacks and multiple methods adapted from vision-language model attacks, and validating them in simulation and real robot environments. The analysis shows that existing attacks mostly cause untargeted failure or action stalling, and targeted attacks capable of driving VLA to execute precise long-horizon actions are largely unexplored. To this end, the paper proposes the targeted backdoor attack BackdoorVLA, which forces the VLA to execute an attacker-specified long-horizon action sequence when a trigger appears, achieving an average targeted success rate of 58.4% in simulation benchmarks and real robot environments, with some tasks reaching 100%. The authors state that this work provides a standardized framework for evaluating VLA vulnerabilities.