Adversarial Attack Detection on Balanced 500-sample non-adaptive
94PrecisionSpectralGuard (Multi-Layer)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| SpectralGuard (Multi-Layer)Setting=Non-adaptive, Attacker=Non-adaptive, Latency=< 15ms/tok, N=500, Model=Mamba-130M2026.03 | 94 | 98 | 96.1 | 0.989 | 6 | |
| Internal L2-Norm MonitorSetting=Non-adaptive, Attacker=Non-adaptive, Latency=< 5ms, N=500, Model=Mamba-130M2026.03 | 62 | 51 | 56 | — | 5 | |
| SpectralGuard (Single-Layer)Setting=Non-adaptive, Attacker=Non-adaptive, Latency=< 5ms/tok, N=500, Model=Mamba-130M2026.03 | 54.9 | 70 | 61.9 | — | 67.2 | |
| Perplexity FilterSetting=Non-adaptive, Attacker=Non-adaptive, Latency=< 1ms, N=500, Model=Mamba-130M2026.03 | 42 | 31 | 36 | — | 2 | |
| Pattern MatcherSetting=Non-adaptive, Attacker=Non-adaptive, Latency=< 1ms, N=500, Model=Mamba-130M2026.03 | 15 | 12 | 13 | — | — | |
| Toxicity ClassifierSetting=Non-adaptive, Attacker=Non-adaptive, Latency=< 1ms, N=500, Model=Mamba-130M2026.03 | 8 | 5 | 6 | — | 1 |