Linguistic Steganography on LLM Steganography Evaluation Dataset
0.201Delta PPLASW (soft)
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| ASW (soft)Language Model=Qwen2.5-7B-Instruct, Bridge context=θbridge trained with forward KL, l_bridge=8, w=102026.04 | 0.201 | 27.6 | 20 | 82.8 | 74.5 | 1.606 | 32.594 | |
| Full contextLanguage Model=Qwen2.5-14B-Instruct, Bridge context=N/A, l_bridge=8, w=102026.04 | 0.612 | 38.9 | 25.7 | 85.3 | 59 | 1.135 | 244.337 | |
| ASW (soft)Language Model=Qwen2.5-14B-Instruct, Bridge context=θbridge trained with forward KL, l_bridge=8, w=102026.04 | 0.761 | 28.8 | 20.2 | 83.3 | 72.5 | 1.536 | 74.986 | |
| ASW (hard)Language Model=Qwen2.5-14B-Instruct, Bridge context=“[CONTEXT TRUNCATED]\n”, l_bridge=8, w=102026.04 | 0.871 | 28.4 | 19.8 | 83 | 74 | 1.529 | 73.446 | |
| Full contextLanguage Model=Qwen2.5-7B-Instruct, Bridge context=N/A, l_bridge=8, w=102026.04 | 1.09 | 40.5 | 25.7 | 85.6 | 56 | 1.024 | 115.43 | |
| ASW (soft)Language Model=Qwen2.5-14B-Instruct, Bridge context=θbridge trained with reverse KL, l_bridge=8, w=102026.04 | 1.903 | 28.6 | 20.3 | 83.1 | 73 | 1.198 | 74.317 | |
| ASW (hard)Language Model=Qwen2.5-7B-Instruct, Bridge context=“[CONTEXT TRUNCATED]\n”, l_bridge=8, w=102026.04 | 3.246 | 24.2 | 19.5 | 81.6 | 83 | 1.012 | 30.868 | |
| ASW (hard)Language Model=Qwen2.5-14B-Instruct, Bridge context=“...\n”, l_bridge=8, w=102026.04 | 3.645 | 23.9 | 19.5 | 81.6 | 82 | 1.095 | 72.203 | |
| ASW (soft)Language Model=Qwen2.5-14B-Instruct, Bridge context=Random untrained θbridge, l_bridge=8, w=102026.04 | 3.948 | 20.9 | 18.9 | 80.1 | 91.5 | 1.029 | 74.604 | |
| ASWLanguage Model=Qwen2.5-14B-Instruct, Bridge context=0 token, l_bridge=8, w=102026.04 | 4.022 | 22.5 | 19.3 | 81.1 | 90 | 1.073 | 71.789 | |
| ASW (soft)Language Model=Qwen2.5-7B-Instruct, Bridge context=Random untrained θbridge, l_bridge=8, w=102026.04 | 4.066 | 20 | 18.5 | 79.7 | 91 | 0.98 | 31.819 | |
| ASW (soft)Language Model=Qwen2.5-7B-Instruct, Bridge context=θbridge trained with reverse KL, l_bridge=8, w=102026.04 | 4.336 | 26.7 | 20 | 82.5 | 77 | 1.124 | 31.499 | |
| ASW (hard)Language Model=Qwen2.5-7B-Instruct, Bridge context=“...\n”, l_bridge=8, w=102026.04 | 4.56 | 24.8 | 19.5 | 81.6 | 81.5 | 1.035 | 30.505 | |
| ASWLanguage Model=Qwen2.5-7B-Instruct, Bridge context=0 token, l_bridge=8, w=102026.04 | 5.157 | 22 | 19 | 80.6 | 88.5 | 0.966 | 29.702 | |
| ASW (soft)Language Model=Qwen2.5-3B-Instruct, Bridge context=θbridge trained with reverse KL, l_bridge=8, w=102026.04 | 12.337 | 25.1 | 19.1 | 81.8 | 79.5 | 1.635 | 14.998 | |
| ASW (soft)Language Model=Qwen2.5-3B-Instruct, Bridge context=Random untrained θbridge, l_bridge=8, w=102026.04 | 12.516 | 20.9 | 18.3 | 80.4 | 90 | 1.348 | 15.286 | |
| ASW (hard)Language Model=Qwen2.5-3B-Instruct, Bridge context=“...\n”, l_bridge=8, w=102026.04 | 12.693 | 23.1 | 18.6 | 81.2 | 85 | 1.601 | 14.528 | |
| ASW (soft)Language Model=Qwen2.5-3B-Instruct, Bridge context=θbridge trained with forward KL, l_bridge=8, w=102026.04 | 13.046 | 25.3 | 19 | 82 | 78.5 | 1.984 | 15.36 | |
| Full contextLanguage Model=Qwen2.5-3B-Instruct, Bridge context=N/A, l_bridge=8, w=102026.04 | 13.316 | 34.5 | 21.3 | 83.9 | 64.5 | 1.604 | 55.951 | |
| ASW (hard)Language Model=Qwen2.5-3B-Instruct, Bridge context=“[CONTEXT TRUNCATED]\n”, l_bridge=8, w=102026.04 | 13.488 | 22 | 18.4 | 80.8 | 85.5 | 1.71 | 14.831 | |
| ASWLanguage Model=Qwen2.5-3B-Instruct, Bridge context=0 token, l_bridge=8, w=102026.04 | 16.369 | 21.6 | 18.4 | 80.7 | 88.5 | 1.431 | 14.367 | |
| WinStegaLanguage Model=Qwen2.5-3B-Instruct, Bridge context=N/A, l_bridge=8, w=102026.04 | 25.743 | 8.3 | 9.9 | 70.1 | 94.5 | 2.588 | 13.121 | |
| WinStegaLanguage Model=Qwen2.5-7B-Instruct, Bridge context=N/A, l_bridge=8, w=102026.04 | 31.486 | 8.2 | 9.8 | 69.9 | 95.5 | 1.995 | 22.25 | |
| WinStegaLanguage Model=Qwen2.5-14B-Instruct, Bridge context=N/A, l_bridge=8, w=102026.04 | 39.483 | 8.2 | 9.9 | 69.8 | 93 | 2.54 | 51.911 |