Linguistic Steganography on databricks-dolly-15k
0.435PPL Change (ΔPPL)Full context
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Full contextBridge context=N/A, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 0.435 | 39.2 | 34 | 85.1 | 60 | 0.96 | 50.517 | |
| ASW (soft)Bridge context=θbridge trained with forward KL, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 1.085 | 21.7 | 21.8 | 82 | 82.5 | 1.534 | 29.921 | |
| ASW (soft)Bridge context=θbridge trained with reverse KL, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 5.084 | 21.4 | 22.5 | 82.5 | 80.5 | 1.047 | 29.941 | |
| ASW (hard)Bridge context="[CONTEXT TRUNCATED]\n", Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 5.165 | 21.4 | 22.6 | 82.3 | 81 | 0.986 | 28.872 | |
| ASW (hard)Bridge context=“...\n”, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 5.244 | 20.8 | 21.9 | 81.7 | 86.5 | 0.933 | 27.874 | |
| ASWBridge context=0 token, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 5.332 | 19.7 | 21.4 | 81.5 | 89.5 | 0.892 | 28.405 | |
| ASW (soft)Bridge context=Random untrained θbridge, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 6.956 | 17.3 | 20.1 | 79.8 | 90.5 | 0.864 | 30.944 | |
| WinStegaBridge context=N/A, Model=Qwen2.5-7B-Instruct, l_bridge=8, w=102026.04 | 35.162 | 6.4 | 9.7 | 66.1 | 96 | 2.996 | 22.549 |