Prompt Injection Attack Detection on Alpaca
100TPRDualSentinel
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DualSentinelBackbone=GPT-4o, Attacks=System, Target Form=OnlyTarget, Sequence Length=Long2026.03 | 100 | 4.8 | |
| PPLModel=LlaMA2 13B, Attacks=Ignore, Target format=OnlyTarget, Target sequence length=Long2026.03 | 100 | 100 | |
| CleanGenModel=LlaMA2 13B, Attacks=Ignore, Target format=OnlyTarget, Target sequence length=Long2026.03 | 100 | 100 | |
| DualSentinelModel=LlaMA2 7B, Attacks=Word, Target format=OnlyTarget, Target sequence length=Long2026.03 | 100 | 11.13 | |
| DualSentinelModel=LlaMA2 7B, Attacks=Phrase, Target format=OnlyTarget, Target sequence length=Long2026.03 | 100 | 10.54 | |
| DualSentinelModel=LlaMA2 7B, Attacks=Sentence, Target format=OnlyTarget, Target sequence length=Long2026.03 | 100 | 9.45 | |
| DualSentinelModel=LlaMA2 7B, Attacks=Joint, Target format=OnlyTarget, Target sequence length=Long2026.03 | 100 | 10.52 | |
| DualSentinelModel=Qwen2.5 7B, Attacks=Word, Target format=OnlyTarget, Target sequence length=Long2026.03 | 100 | 7.01 | |
| DualSentinelModel=Qwen2.5 7B, Attacks=Phrase, Target format=OnlyTarget, Target sequence length=Long2026.03 | 100 | 6.75 | |
| DualSentinelModel=Qwen2.5 7B, Attacks=Sentence, Target format=OnlyTarget, Target sequence length=Long2026.03 | 100 | 5.99 | |
| DualSentinelModel=Qwen2.5 7B, Attacks=Joint, Target format=OnlyTarget, Target sequence length=Long2026.03 | 99.89 | 8.08 | |
| DualSentinelModel=Qwen2.5 14B, Attacks=Complete, Target format=OnlyTarget, Target sequence length=Long2026.03 | 99 | 3 | |
| DualSentinelBackbone=GPT-4o, Attacks=Average, Target Form=OnlyTarget, Sequence Length=Long2026.03 | 97.99 | 4.8 | |
| DualSentinelModel=LlaMA2 13B, Attacks=System, Target format=OnlyTarget, Target sequence length=Long2026.03 | 97.39 | 16 | |
| DualSentinelBackbone=GPT-4o-mini, Attacks=System, Target Form=OnlyTarget, Sequence Length=Long2026.03 | 97.12 | 3.6 | |
| ConfGuardModel=LlaMA2 13B, Attacks=Ignore, Target format=OnlyTarget, Target sequence length=Long2026.03 | 96 | 31.7 | |
| DualSentinelModel=Qwen2.5 14B, Attacks=Ignore, Target format=OnlyTarget, Target sequence length=Long2026.03 | 96 | 3 | |
| DualSentinelBackbone=GPT-4o, Attacks=Complete, Target Form=OnlyTarget, Sequence Length=Long2026.03 | 95.98 | 4.8 | |
| DualSentinelModel=LlaMA2 13B, Attacks=Complete, Target format=OnlyTarget, Target sequence length=Long2026.03 | 94.8 | 16 | |
| DualSentinelModel=LlaMA2 13B, Attacks=Tree, Target format=OnlyTarget, Target sequence length=Long2026.03 | 94.04 | 16 | |
| DualSentinelBackbone=GPT-4o-mini, Attacks=Average, Target Form=OnlyTarget, Sequence Length=Long2026.03 | 92.16 | 3.6 | |
| DualSentinelModel=Qwen2.5 14B, Attacks=System, Target format=OnlyTarget, Target sequence length=Long2026.03 | 91 | 3 | |
| DualSentinelBackbone=GPT-4o-mini, Attacks=Complete, Target Form=OnlyTarget, Sequence Length=Long2026.03 | 87.2 | 3.6 | |
| DualSentinelModel=LlaMA2 13B, Attacks=Ignore, Target format=OnlyTarget, Target sequence length=Long2026.03 | 87.08 | 16 | |
| DualSentinelModel=Qwen2.5 14B, Attacks=Tree, Target format=OnlyTarget, Target sequence length=Long2026.03 | 86 | 3 | |
| ONIONModel=LlaMA2 13B, Attacks=Ignore, Target format=OnlyTarget, Target sequence length=Long2026.03 | 56.6 | 67.1 | |
| STRIPModel=LlaMA2 13B, Attacks=Ignore, Target format=OnlyTarget, Target sequence length=Long2026.03 | 49.92 | 74 | |
| ParagraphModel=LlaMA2 13B, Attacks=Ignore, Target format=OnlyTarget, Target sequence length=Long2026.03 | 38.26 | 6 |