New SentAttack method targets dense retrieval models in AI systems

Researchers have developed SentAttack, a novel sentence-level black-box adversarial attack method specifically designed for dense retrieval (DR) models within retrieval-augmented generation (RAG) systems. This method addresses the limitations of existing word-level attacks by generating low-ranked, irrelevant documents that can mislead DR models. SentAttack operates in two stages: first, it trains a surrogate DR model using data from the black-box RAG system, and second, it uses this surrogate model to create adversarial documents by concatenating sentence-level clusters with the target query. AI

IMPACT This research could lead to more robust dense retrieval models by highlighting vulnerabilities to sentence-level adversarial attacks.

RANK_REASON The item is an academic paper detailing a new method for attacking AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.IR (Information Retrieval) →

• alphaXiv
• CatalyzeX
• DagsHub
• Dense Passage Retrieval for Open-Domain Question Answering
• Gotit.pub
• Hugging Face
• network management
• neural ranking model
• retrieval-augmented generation
• ScienceCast
• SentAttack

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →