Jailbreak Defense on X-Teaming
0Attack Success Rate (ASR)Prompt Guard
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Prompt GuardLatency (Sec.)=22.67, Memory Overhead (MB)=1864.112026.04 | 0 | 0 | |
| SelfGraderLatency (Sec.)=1.24, Memory Overhead (MB)=479.812026.04 | 0 | 0 | |
| GradientCuffLatency (Sec.)=26.57, Memory Overhead (MB)=2063.762026.04 | 1 | 1 | |
| SelfDefend (Intent)Latency (Sec.)=3.11, Memory Overhead (MB)=13457.402026.04 | 22 | 22 | |
| SelfDefend (Direct)Latency (Sec.)=1.19, Memory Overhead (MB)=13452.932026.04 | 59 | 61 | |
| GuardReasoner (Pre)Latency (Sec.)=14.87, Memory Overhead (MB)=15617.332026.04 | 64 | 65 | |
| GuardReasoner (Post)Latency (Sec.)=16.71, Memory Overhead (MB)=15618.132026.04 | 76 | 83 | |
| Llama Guard (Pre)Latency (Sec.)=0.96, Memory Overhead (MB)=13910.612026.04 | 77 | 81 | |
| Llama Guard (Post)Latency (Sec.)=1.17, Memory Overhead (MB)=13910.612026.04 | 85 | 91 | |
| LLama-3-8B-Instruct (No Defense)Latency (Sec.)=1.472026.04 | 91 | — | |
| Perplexity FilterLatency (Sec.)=0.49, Memory Overhead (MB)=13571.942026.04 | 91 | 100 |