Linguistic Steganography on Super-NaturalInstructions
0.669ΔPPLFull context
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Full contextBridge context=N/A, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 0.669 | 33.3 | 34.2 | 83.9 | 72 | 0.819 | 13.79 | |
| ASW (soft)Bridge context=θbridge trained with forward KL, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 5.786 | 10.9 | 16.5 | 78.9 | 86 | 1.75 | 38.24 | |
| ASW (soft)Bridge context=θbridge trained with reverse KL, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 10.82 | 10.9 | 16.7 | 79 | 85.5 | 1.211 | 38.652 | |
| ASWBridge context=0 token, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 11.507 | 10.5 | 16.2 | 78.6 | 92.5 | 0.98 | 32.34 | |
| ASW (hard)Bridge context=“...\n”, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 12.117 | 10.7 | 16.3 | 78.8 | 87.5 | 0.995 | 32.654 | |
| ASW (soft)Bridge context=Random untrained θbridge, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 12.23 | 10.6 | 16.2 | 78.4 | 93 | 1.03 | 40.115 | |
| ASW (hard)Bridge context="[CONTEXT TRUNCATED]\n", Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 12.622 | 10.7 | 16.4 | 79.2 | 88 | 1.03 | 37.687 | |
| WinStegaBridge context=N/A, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 22.171 | 2.4 | 7.7 | 64.4 | 98.5 | 3.003 | 22.519 |