Adversarial Attack Detection on Held-out n = 560 (20% stratified test)
89.1Detection RateCombined
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| CombinedSystem=Dual-layer defense framework2026.06 | 89.1 | 0.3 | 99.7 | 94.2 | 89.1 | 99.7 | 0.923 | 0.974 | 94.3 | |
| Prompt LayerType=Aggregate Layer2026.06 | 73.3 | 0 | 100 | 84.6 | — | — | — | — | — | |
| Linguistic PatternsLayer=Prompt Layer2026.06 | 68.2 | 0 | 100 | 81.1 | — | — | — | — | — | |
| Code LayerType=Aggregate Layer2026.06 | 67.8 | 0.3 | 99.6 | 80.7 | — | — | — | — | — | |
| Position MonitoringLayer=Prompt Layer2026.06 | 62.9 | 0 | 100 | 77.2 | — | — | — | — | — | |
| Semantic ValidationLayer=Prompt Layer2026.06 | 54.1 | 0 | 100 | 70.3 | — | — | — | — | — | |
| DifferentialLayer=Code Layer2026.06 | 51.3 | 0 | 100 | 67.7 | — | — | — | — | — | |
| Static AnalysisLayer=Code Layer2026.06 | 42.1 | 0 | 100 | 59.3 | — | — | — | — | — | |
| Perplexity AnalysisLayer=Prompt Layer2026.06 | 41.7 | 2.3 | 94.7 | 58 | — | — | — | — | — | |
| Pattern MatchingLayer=Code Layer2026.06 | 38.7 | 0 | 100 | 55.8 | — | — | — | — | — | |
| Taint AnalysisLayer=Code Layer2026.06 | 29.4 | 1.2 | 96.1 | 44.9 | — | — | — | — | — |