Research map / Adversarial robustness and security / Attacks on text and graphs
Through the Stealth Lens: Attention-Aware Defenses Against Poisoning in RAG
Sarthak Choudhary, Nils Palumbo, Ashish Hooda, Krishnamurthy Dj Dvijotham, Somesh Jha
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 AV Filter to vanilla RAG | RAG pipeline · RQA-MC clean (avg over LLMs) | clean accuracy: 61.4 -> 59.3 | Hurt |
| add AV Filter to vanilla RAG | RAG pipeline · RQA-MC under attack (avg) | robust accuracy: 44.4 -> 55.7 | Improved |
| add AV Filter to vanilla RAG | RAG pipeline · RQA-MC under attack (avg) | attack success rate: 88.2 -> 6.6 | Improved |
| use AV Filter instead of Keyword defense | RAG pipeline · RQA-MC under attack (avg) | robust accuracy: 53.9 -> 55.7 | Improved |
| use AV Filter instead of Keyword defense | RAG pipeline · RQA-MC under attack (avg) | attack success rate: 6.1 -> 6.6 | No clear effect |
| use AV Filter instead of Decoding defense | RAG pipeline · RQA-MC clean (avg over LLMs) | clean accuracy: 50.3 -> 59.3 | Improved |
| combine AV Filter with Keyword | Keyword defense · RACC and ASR across settings | robust accuracy and ASR: 18/18 cases | Improved |
| add adaptive attention-aware attack against AV Filter | AV Filter · RQA-MC | attack success rate: 6.6 -> 20 | Hurt |
| use AV Filter instead of Keyword on multi-hop QA | RAG pipeline · HotpotQA | attack success rate | Improved |
| vary NormScore alpha from 5 to infinity (ablation) | AV Filter · RQA-MC clean (Mistral-7B) | clean accuracy: 73.0 -> 76.0 | Improved |
| vary NormScore alpha from 5 to infinity (ablation) | AV Filter · RQA-MC clean (Llama2-C) | clean accuracy: 79.0 -> 75.0 | Hurt |
| use AV Filter instead of Keyword on clean Llama2-C | RAG pipeline · RQA-MC clean (Llama2-C) | clean accuracy: 56.0 -> 79.0 | Improved |
Closest papers by meaning
- A Set of Generalized Components to Achieve Effective Poison-only Clean-label Backdoor Attacks with Collaborative Sample Selection and Triggers (NeurIPS 2025 · similarity 0.84)
- Shadowcast: Stealthy Data Poisoning Attacks Against Vision-Language Models (NeurIPS 2024 · similarity 0.84)
- Activation Gradient based Poisoned Sample Detection Against Backdoor Attacks (ICLR 2025 · similarity 0.83)
- Demystifying Poisoning Backdoor Attacks from a Statistical Perspective (ICLR 2024 · similarity 0.83)
- MemPoison: Uncovering Persistent Memory Threats and Structural Blind Spots in LLM Agents (NeurIPS 2026 · similarity 0.83)
- How does information access affect LLM monitors' ability to detect sabotage? (ICML 2026 · similarity 0.83)
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.
