Research map / Adversarial robustness and security / Attacks on text and graphs
*MemPot*: Defend Against Memory Extraction Attack with Optimized Honeypots
Yuhao Wang, Shengfang Zhai, Guanghao Jin, Yinpeng Dong, Linyi Yang, Jiaheng Zhang
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 |
|---|---|---|---|
| inject optimized honeypots with sequential detection vs optimal static sequential detector | MemPot · EHRAgent, MEXTRA-Cosine | AUROC: 0.70 -> 0.99 | Improved |
| inject optimized honeypots (vs Optimal-Seq without honeypots) | MemPot · EHRAgent, MEXTRA-Cosine | first detection time (turns): 22 -> 2 | Improved |
| use MemPot instead of ControlNet / agent detectors | MemPot · IKEA, HealthMagicCare | AUROC: 0.46 -> 0.96 | Improved |
| add honeypots to memory | MemPot · RAP WebShop utility | success rate: 44.6 -> 43.8 | No clear effect |
| add honeypots to memory | MemPot · RAP WebShop utility | score: 67.1 -> 65.4 | Hurt |
| add honeypots to memory | MemPot · Pokemon QA and MCQ | Rouge-L: 0.67 -> 0.66 | No clear effect |
| use similarity-ratio LLR instead of count-ratio LLR (ablation) | MemPot · EHRAgent, MEXTRA-Cosine | AUROC: 0.99 -> 0.74 | Hurt |
| use global-similarity LLR instead of count LLR (ablation) | MemPot · EHRAgent, MEXTRA-Cosine | first detection time (turns): 2 -> 1 | Improved |
| increase honeypot ratio from 0.6% to 4% (ablation) | MemPot · RAP WebShop, MEXTRA-Cosine | TPR@1%FPR: 0.77 -> 0.94 | Improved |
| increase honeypot ratio from 0.6% to 4% (ablation) | MemPot · RAP WebShop, MEXTRA-Cosine | first detection time (turns): 7 -> 2 | Improved |
| increase honeypot ratio from 0.6% to 4% (ablation) | MemPot · RAP WebShop, MEXTRA-Cosine | AUROC: 1.00 -> 1.00 | No clear effect |
| stack new honeypots for new records instead of re-optimizing end to end (ablation) | MemPot · EHRAgent, MEXTRA-General (500 records) | AUROC: 0.94 -> 0.98 | No clear effect |
Closest papers by meaning
- A-MemGuard: A Proactive Defense Framework For LLM-Based Agent Memory (ICML 2026 · similarity 0.90)
- MemPoison: Uncovering Persistent Memory Threats and Structural Blind Spots in LLM Agents (NeurIPS 2026 · similarity 0.88)
- Memory Injection Attacks on LLM Agents via Query-Only Interaction (NeurIPS 2025 · similarity 0.86)
- Speculative Safety Honeypot: Toward Proactive Defense Against Multi-turn Agent Attacks (ICML 2026 · similarity 0.85)
- RedVisor: Reasoning-Aware Prompt Injection Defense via Zero-Copy KV Cache Reuse (ICML 2026 · similarity 0.85)
- Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models (ICLR 2025 · similarity 0.85)
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.
