Researchers have developed AdvNav, a novel black-box adversarial attack framework designed to disrupt Vision-and-Language Navigation (VLN) systems. Unlike previous methods requiring white-box access, AdvNav operates solely on observable inputs and outputs, making it more applicable to real-world deployed systems. The framework uses a dual-granularity behavior-based feedback mechanism to guide its optimization strategy, effectively discovering noise configurations that degrade navigation performance. Evaluations showed AdvNav achieved high attack success rates against transformer-based and LLM-based VLN models on the R2R dataset, highlighting critical perception vulnerabilities. AI
IMPACT This research highlights critical vulnerabilities in current vision-language navigation systems, potentially guiding the development of more robust and secure AI agents.
RANK_REASON The item is an academic paper detailing a new adversarial attack method for AI systems. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
