Research map / Adversarial robustness and security / Attacks on text and graphs
RADAR: Defending RAG Dynamically against Retrieval Corruption
Ziyuan Chen, Yueming Lyu, Yi Liu, Weixiang Han, Jing Dong, Caifeng Shan, Tieniu Tan
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 |
|---|---|---|---|
| use RADAR instead of vanilla RAG | DeepSeek RAG · RQA PIA Pos 1 top-k=10 | accuracy (%): 25.0 -> 69.0 | Improved |
| use RADAR instead of vanilla RAG | DeepSeek RAG · Dynamic PIA Pos 1 top-k=50 | accuracy (%): 12.41 -> 63.60 | Improved |
| use RADAR instead of vanilla RAG in benign dynamic setting | DeepSeek RAG · Dynamic benign cumulative | accuracy (%): 70.76 -> 74.02 | Improved |
| use RADAR instead of vanilla RAG | DeepSeek RAG · Bio PIA Pos 50 top-k=50 | factual accuracy: 10.2 -> 80.6 | Improved |
| use RADAR instead of ReliabilityRAG | DeepSeek RAG · Dynamic PIA Pos 1 | accuracy (%): 53.61 -> 63.60 | Improved |
| use RADAR instead of ReliabilityRAG | DeepSeek RAG · NQ PIA Pos 1 top-k=10 | accuracy (%): 65.2 -> 63.4 | Hurt |
| use RADAR instead of ReliabilityRAG (runtime) | DeepSeek RAG · RQA top-k=50 | total time (s): 43 -> 87 | Hurt |
| swap NLI model DeBERTa-v3 to BART or ModernBERT (ablation) | RADAR · RQA PIA top-k=50 Pos 1 | ASR (%): 5.0 -> 7.0 | No clear effect |
| randomly perturb NLI outputs at rate 0.5 (ablation) | RADAR · RQA PIA top-k=10 Pos 1 | ASR (%): 11.0 -> 18.0 | Hurt |
| adaptive attack versus non-adaptive attack | RADAR · RQA PIA top-k=50 Pos 1 | ASR (%): 5.0 -> 8.0 | Hurt |
| vary post-processing threshold lambda 0.1 to 0.5 (ablation) | RADAR · RQA PIA | accuracy and ASR | No clear effect |
| use lightweight history instead of cumulative snapshot | Vanilla RAG · Dynamic PIA Pos 1 | accuracy (%): 12.41 -> 37.68 | Improved |
Closest papers by meaning
- ReliabilityRAG: Effective and Provably Robust Defense for RAG-based Web-Search (NeurIPS 2025 · similarity 0.88)
- Worse than Zero-shot? A Fact-Checking Dataset for Evaluating the Robustness of RAG Against Misleading Retrievals (NeurIPS 2025 · similarity 0.83)
- Follow My Instruction and Spill the Beans: Scalable Data Extraction from Retrieval-Augmented Generation Systems (ICLR 2025 · similarity 0.82)
- Closing the Gap: Achieving Better Accuracy-Robustness Tradeoffs against Query-Based Attacks (AAAI 2024 · similarity 0.82)
- On the Vulnerability of Applying Retrieval-Augmented Generation within Knowledge-Intensive Application Domains (ICML 2025 · similarity 0.81)
- LT-Defense: Searching-free Backdoor Defense via Exploiting the Long-tailed Effect (NeurIPS 2024 · similarity 0.81)
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.
