Global Source Attribution on MLAAD Few-Shot Open-Set (test)
84.52F1-Macro (All vs All)Strategy 2.B (Residual Modeling)
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Strategy 2.B (Residual Modeling)MSE (λ)=0.05, K=52026.07 | 84.52 | 90.6 | 81.41 | 9.19 | |
| Strategy 2.A (Factorized Prototypes)MSE (λ)=0.0, K=52026.07 | 83.97 | 90.05 | 80.86 | 9.19 | |
| Strategy 2.A (Factorized Prototypes)MSE (λ)=0.05, K=52026.07 | 83.67 | 91.34 | 79.75 | 11.59 | |
| Strategy 1 (Orthonormal Prototypes)MSE (λ)=0.0, K=52026.07 | 82.7 | 90.68 | 78.61 | 12.07 | |
| Strategy 1 (Orthonormal Prototypes)MSE (λ)=0.05, K=52026.07 | 82.56 | 91.14 | 78.17 | 12.97 | |
| Strategy 1 (Orthonormal Prototypes)MSE (λ)=0.1, K=52026.07 | 82.55 | 91.22 | 78.11 | 13.11 | |
| Strategy 2.B (Residual Modeling)MSE (λ)=0.05, K=32026.07 | 82.12 | — | — | — | |
| Strategy 1 (Orthonormal Prototypes)MSE (λ)=1.0, K=52026.07 | 81.44 | 91.39 | 76.35 | 15.04 | |
| Strategy 2.A (Factorized Prototypes)MSE (λ)=0.0, K=32026.07 | 81.44 | — | — | — | |
| Baseline (ArcFace)MSE (λ)=–, K=52026.07 | 80.11 | 89.17 | 75.47 | 13.7 | |
| Strategy 1 (Orthonormal Prototypes)MSE (λ)=0.0, K=32026.07 | 80.09 | — | — | — | |
| Baseline (ArcFace)MSE (λ)=–, K=32026.07 | 77.49 | — | — | — | |
| Strategy 2.B (Residual Modeling)MSE (λ)=0.05, K=12026.07 | 72.11 | — | — | — | |
| Strategy 2.A (Factorized Prototypes)MSE (λ)=0.0, K=12026.07 | 70.94 | — | — | — | |
| Strategy 1 (Orthonormal Prototypes)MSE (λ)=0.0, K=12026.07 | 70.18 | — | — | — | |
| Baseline (ArcFace)MSE (λ)=–, K=12026.07 | 68.01 | — | — | — |