Reproduce “ATAAT: Adaptive Threat-Aware Adversarial Tuning Framework against Backdoor Attacks on Vision-Language-Action Models” on OpenResearch

About this paper

Kewei Chen, Yayu Long, Shuai Li, Mingsheng Shang

Researchers from the Chinese Academy of Sciences and the University of Oulu introduced the Adaptive Threat-Aware Adversarial Tuning (ATAAT) framework, designed to overcome "Gradient Interference" in Vision-Language-Action (VLA) backdoor attacks. This framework enables robust backdoor injection in VLA models, achieving over 80% targeted attack success rates with only 5% poisoning, while preserving high benign task performance and enhancing semantic robustness against linguistic variations.

Key insights

• "Gradient Interference" is identified as the core obstacle to robust backdoor attacks in VLA models, causing conflicting optimization directions quantified by consistently negative gradient cosine similarity.
• Orthogonalizing the optimization gradients for benign and backdoor tasks, either implicitly through data poisoning or explicitly through parameter isolation, is crucial for enabling effective and stealthy backdoor injection.
• Backdoor logic can be robustly bound to abstract semantic concepts (e.g., object states, human attributes) rather than superficial pixel patterns, enhancing attack stealthiness and resilience against linguistic and environmental variations.

Summary by alphaXiv.