Linguistic Steganography on InstructionWild
0.201ΔPPLASW (soft)
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| ASW (soft)Bridge context=θbridge trained with forward KL, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 0.201 | 27.6 | 20 | 82.8 | 74.5 | 1.606 | 32.594 | |
| Full contextBridge context=N/A, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 1.09 | 40.5 | 25.7 | 85.6 | 56 | 1.024 | 115.43 | |
| ASW (hard)Bridge context="[CONTEXT TRUNCATED]\n", Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 3.246 | 24.2 | 19.5 | 81.6 | 83 | 1.012 | 30.868 | |
| ASW (soft)Bridge context=Random untrained θbridge, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 4.066 | 20 | 18.5 | 79.7 | 91 | 0.98 | 31.819 | |
| ASW (soft)Bridge context=θbridge trained with reverse KL, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 4.336 | 26.7 | 20 | 82.5 | 77 | 1.124 | 31.499 | |
| ASW (hard)Bridge context=“...\n”, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 4.56 | 24.8 | 19.5 | 81.6 | 81.5 | 1.035 | 30.505 | |
| ASWBridge context=0 token, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 5.157 | 22 | 19 | 80.6 | 88.5 | 0.966 | 29.702 | |
| WinStegaBridge context=N/A, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 31.486 | 8.2 | 9.8 | 69.9 | 95.5 | 1.995 | 22.25 |