Research map / Adversarial robustness and security / Black-box and transfer attacks
``Someone Hid It!'': Query-Agnostic Black-Box Attacks on LLM-Based Retrieval
Jiate Li, Defu Cao, Li Li, Wei Yang, Yuehan Qin, Chenxiao Yu, Tiannuo Yang, Ryan A. Rossi, Yan Liu, Xiyang Hu, Yue Zhao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add word-embedding surrogate instead of last-hidden (ablation) | DQ-A · BRIGHT Economics, Qwen1.5 victim | Recall@25 after attack (lower is better): 0.289 -> 0.223 | Improved |
| add word-embedding surrogate instead of last-hidden (ablation) | DQ-A · BRIGHT Robotics, jina victim | Recall@25 after attack (lower is better): 0.245 -> 0.253 | Hurt |
| add adversarial query-document learning over plain sampled-query GCG (ablation) | DQ-A-H vs S-GCG · BRIGHT Economics, Qwen1.5 victim | Recall@25 after attack (lower is better): 0.299 -> 0.289 | Improved |
| add adversarial learning plus word-embedding over S-GCG (ablation) | DQ-A · BRIGHT Robotics, Qwen1.5 victim | Recall@25 after attack (lower is better): 0.181 -> 0.152 | Improved |
| increase sampled query count from 5 to 10 (ablation) | DQ-A · BRIGHT, all retrievers | attack-induced performance drop | Improved |
| increase sampled query count from 10 to 15 (ablation) | DQ-A · BRIGHT, all retrievers | attack-induced performance drop | Mixed |
| increase injected token amount (ablation) | DQ-A · BRIGHT, all retrievers | attack-induced performance drop | No clear effect |
| inject tokens at the middle instead of beginning or end (ablation) | DQ-A · BRIGHT, Embedding-Gemma | attack-induced performance drop | Hurt |
| add a reranker after the retriever | DQ-A · BRIGHT Economics, BGE-M3 + BGE-reranker-v2-M3 | NDCG@25: 0.139 -> 0.131 | No clear effect |
| reverse the attack losses to boost a document | DQ-A reversed · BRIGHT Economics, single picked document | Recall@50: 0.00227 -> 0.00601 | Improved |
| use Qwen3-Embedding-0.6B as victim | DQ-A · BRIGHT average | Recall@25 after attack: 0.270 -> 0.264 | No clear effect |
Closest papers by meaning
- On the Vulnerability of Applying Retrieval-Augmented Generation within Knowledge-Intensive Application Domains (ICML 2025 · similarity 0.89)
- Adversarial Search Engine Optimization for Large Language Models (ICLR 2025 · similarity 0.88)
- Soft Prompt Threats: Attacking Safety Alignment and Unlearning in Open-Source LLMs through the Embedding Space (NeurIPS 2024 · similarity 0.88)
- Bits Leaked per Query: Information-Theoretic Bounds for Adversarial Attacks on LLMs (NeurIPS 2025 · similarity 0.88)
- Can We Infer Confidential Properties of Training Data from LLMs? (NeurIPS 2025 · similarity 0.88)
- Exploiting LLM Quantization (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
