Researchers have developed AdvNav, a novel black-box adversarial attack framework designed to disrupt Vision-Language Navigation (VLN) systems. Unlike previous methods requiring white-box access or focusing on single-step tasks, AdvNav operates without model gradients and targets the sequential perception-action loop. It utilizes a dual-granularity behavior-based feedback mechanism, incorporating trajectory and action-level performance scores, to guide an optimization strategy that iteratively discovers disruptive noise configurations. Evaluations showed AdvNav achieved high attack success rates against transformer-based and LLM-based VLN models on the R2R dataset, highlighting significant vulnerabilities in current systems. AI
IMPACT Highlights critical perception vulnerabilities in VLN models, potentially driving research into more resilient AI navigation systems.
RANK_REASON Academic paper detailing a new adversarial attack method on AI systems. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
