Defense effectiveness against memory-mediated attacks on ShopBench-Agent
8.3ASR-MCognitiveGuard
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| CognitiveGuardk (number of memories)=52026.04 | 8.3 | 9.7 | 0.8 | 0.89 | 0.3 | 6.5 | |
| CG-ParallelParallelization=parallel queries in the parallel setting2026.04 | 8.3 | 9.7 | 0.8 | 0.89 | 0.3 | 1.8 | |
| CG-LiteBackbone=LLaMA-8B, k (number of memories)=32026.04 | 12.1 | 13.8 | 1.1 | 0.84 | 0.3 | 1.5 | |
| DiffPure2026.04 | 45.2 | 48.7 | 2.1 | 0.52 | 0.3 | — | |
| TrustRAGModality Adaptation=Adapted from text-only settings2026.04 | 48.7 | 51.3 | 2.3 | 0.51 | — | 0.4 | |
| Ret. Adv. Train2026.04 | 52.1 | 55.4 | 2.5 | 0.48 | — | 0.05 | |
| Input Filter2026.04 | 71.5 | 74.2 | 3.8 | 0.38 | — | 0.1 | |
| Output Mod.2026.04 | 78.9 | 81.3 | 4 | 0.34 | — | 0.2 | |
| ProvenanceModality Adaptation=Adapted from text-only settings2026.04 | 79.8 | 82.4 | 4.1 | 0.33 | — | 0.08 | |
| No Defense2026.04 | 82.3 | 85.1 | 4.2 | 0.31 | — | — |