Closed-world Model Attribution on FFHQ near-checkpoint benchmark (StyleGAN2)
98.83Clean AccuracySPRINT
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| SPRINTSignal=Learned, Gradient derivation method=Differentiable WBA2025.08 | 98.83 | 0.83 | 0.22 | 0.94 | 0.06 | 1.06 | 0.07 | 0.21 | 2.35 | |
| Wang20Signal=Learned, Gradient derivation method=Differentiable WBA2025.08 | 98.42 | 100 | 100 | 85.5 | 54.39 | 7.61 | 8.67 | 12 | 6.17 | |
| Song24-FreqSignal=Frequency, Gradient derivation method=WBA through a learned surrogate extractor2025.08 | 78.33 | 54.51 | 15.29 | 69.41 | 15.69 | 35.29 | 40 | 35.29 | 58.82 | |
| Durall20Signal=Frequency, Gradient derivation method=WBA through a learned surrogate extractor2025.08 | 78.15 | 92.29 | 26.63 | 80.9 | 18.92 | 16.8 | 69.97 | 25.62 | 71.9 | |
| Giudice21Signal=Frequency, Gradient derivation method=Differentiable WBA2025.08 | 62.95 | 100 | 99.94 | 85.78 | 24.17 | 14.06 | 63.33 | 63.5 | 78 | |
| Song24-RGBSignal=RGB, Gradient derivation method=WBA through a learned surrogate extractor2025.08 | 59.22 | 41.76 | 16.02 | 3.06 | 0.76 | 1.21 | 0.78 | 0.4 | 1 | |
| Nataraj19Signal=RGB, Gradient derivation method=WBA through a learned surrogate extractor2025.08 | 58.48 | 84.11 | 17.72 | 86 | 22.44 | 64.11 | 57.17 | 57.67 | 67 | |
| Song24-SLSignal=Learned, Gradient derivation method=WBA through a learned surrogate extractor2025.08 | 57.3 | 76.49 | 19.9 | 74.5 | 19.11 | 5.28 | 26.83 | 15.5 | 38.67 |