Model Inversion on CelebA (private)
93.52AccuracyNo Def.
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| No Def.Attack Method=LOMMA-K, D_pub=CelebA, D_pretrain=MS-CelebA-1M, Target Architecture=IR152, Parameter Ratio (|θc|/|θt|)=62.6/62.62024.05 | 93.52 | 96.4 | 99.67 | 1,038 | |
| No Def.Attack Method=LOMMA-K, D_pub=FFHQ, D_pretrain=MS-CelebA-1M, Target Architecture=IR152, Parameter Ratio (|θc|/|θt|)=62.6/62.62024.05 | 93.52 | 73.47 | 90 | 1,290 | |
| No Def.Attack Method=LOMMA-G, D_pub=CelebA, D_pretrain=MS-CelebA-1M, Target Architecture=IR152, Parameter Ratio (|θc|/|θt|)=62.6/62.62024.05 | 93.52 | 64.67 | 86 | 1,401 | |
| No Def.Attack Method=LOMMA-G, D_pub=FFHQ, D_pretrain=MS-CelebA-1M, Target Architecture=IR152, Parameter Ratio (|θc|/|θt|)=62.6/62.62024.05 | 93.52 | 45.2 | 70.67 | 1,503 | |
| No Def.Attack Method=LOMMA-K, D_pub=CelebA, D_pretrain=ImageNet1K, Target Architecture=VGG16, Parameter Ratio (|θc|/|θt|)=16.8/16.82024.05 | 89 | 95.67 | 96.68 | 1,158 | |
| No Def.Attack Method=LOMMA-K, D_pub=FFHQ, D_pretrain=ImageNet1K, Target Architecture=VGG16, Parameter Ratio (|θc|/|θt|)=16.8/16.82024.05 | 89 | 58.6 | 86 | 1,390 | |
| No Def.Attack Method=LOMMA-G, D_pub=CelebA, D_pretrain=ImageNet1K, Target Architecture=VGG16, Parameter Ratio (|θc|/|θt|)=16.8/16.82024.05 | 89 | 56 | 79 | 1,454 | |
| No Def.Attack Method=LOMMA-G, D_pub=FFHQ, D_pretrain=ImageNet1K, Target Architecture=VGG16, Parameter Ratio (|θc|/|θt|)=16.8/16.82024.05 | 89 | 27 | 52.33 | 1,642 | |
| No Def.Attack Method=LOMMA-K, D_pub=CelebA, D_pretrain=MS-CelebA-1M, Target Architecture=FaceNet64, Parameter Ratio (|θc|/|θt|)=35.4/35.42024.05 | 88.5 | 89.33 | 98.67 | 1,226 | |
| No Def.Attack Method=LOMMA-K, D_pub=FFHQ, D_pretrain=MS-CelebA-1M, Target Architecture=FaceNet64, Parameter Ratio (|θc|/|θt|)=35.4/35.42024.05 | 88.5 | 70.27 | 90.33 | 1,391 | |
| No Def.Attack Method=LOMMA-G, D_pub=CelebA, D_pretrain=MS-CelebA-1M, Target Architecture=FaceNet64, Parameter Ratio (|θc|/|θt|)=35.4/35.42024.05 | 88.5 | 60 | 80 | 1,501 | |
| No Def.Attack Method=LOMMA-G, D_pub=FFHQ, D_pretrain=MS-CelebA-1M, Target Architecture=FaceNet64, Parameter Ratio (|θc|/|θt|)=35.4/35.42024.05 | 88.5 | 30.6 | 62 | 1,625 | |
| TL-DMIAttack Method=LOMMA-K, D_pub=CelebA, D_pretrain=MS-CelebA-1M, Target Architecture=IR152, Parameter Ratio (|θc|/|θt|)=17.8/62.62024.05 | 86.7 | 77.73 | 94.67 | 1,305 | |
| TL-DMIAttack Method=LOMMA-K, D_pub=FFHQ, D_pretrain=MS-CelebA-1M, Target Architecture=IR152, Parameter Ratio (|θc|/|θt|)=17.8/62.62024.05 | 86.7 | 45.27 | 74.33 | 1,474 | |
| TL-DMIAttack Method=LOMMA-G, D_pub=CelebA, D_pretrain=MS-CelebA-1M, Target Architecture=IR152, Parameter Ratio (|θc|/|θt|)=17.8/62.62024.05 | 86.7 | 41.87 | 70.67 | 1,551 | |
| TL-DMIAttack Method=LOMMA-G, D_pub=FFHQ, D_pretrain=MS-CelebA-1M, Target Architecture=IR152, Parameter Ratio (|θc|/|θt|)=17.8/62.62024.05 | 86.7 | 22.87 | 43.67 | 1,650 | |
| TL-DMIAttack Method=LOMMA-K, D_pub=CelebA, D_pretrain=ImageNet1K, Target Architecture=VGG16, Parameter Ratio (|θc|/|θt|)=13.9/16.82024.05 | 83.41 | 75.67 | 91.68 | 1,304 | |
| TL-DMIAttack Method=LOMMA-K, D_pub=CelebA, D_pretrain=MS-CelebA-1M, Target Architecture=FaceNet64, Parameter Ratio (|θc|/|θt|)=34.4/35.42024.05 | 83.41 | 79.6 | 97 | 1,345 | |
| TL-DMIAttack Method=LOMMA-K, D_pub=FFHQ, D_pretrain=ImageNet1K, Target Architecture=VGG16, Parameter Ratio (|θc|/|θt|)=13.9/16.82024.05 | 83.41 | 36 | 65 | 1,550 | |
| TL-DMIAttack Method=LOMMA-K, D_pub=FFHQ, D_pretrain=MS-CelebA-1M, Target Architecture=FaceNet64, Parameter Ratio (|θc|/|θt|)=34.4/35.42024.05 | 83.41 | 19.53 | 41.33 | 1,759 | |
| TL-DMIAttack Method=LOMMA-G, D_pub=CelebA, D_pretrain=ImageNet1K, Target Architecture=VGG16, Parameter Ratio (|θc|/|θt|)=13.9/16.82024.05 | 83.41 | 22 | 45.33 | 1,709 | |
| TL-DMIAttack Method=LOMMA-G, D_pub=CelebA, D_pretrain=MS-CelebA-1M, Target Architecture=FaceNet64, Parameter Ratio (|θc|/|θt|)=34.4/35.42024.05 | 83.41 | 43.67 | 65 | 1,616 | |
| TL-DMIAttack Method=LOMMA-G, D_pub=FFHQ, D_pretrain=ImageNet1K, Target Architecture=VGG16, Parameter Ratio (|θc|/|θt|)=13.9/16.82024.05 | 83.41 | 8.87 | 24 | 1,829 | |
| TL-DMIAttack Method=LOMMA-G, D_pub=FFHQ, D_pretrain=MS-CelebA-1M, Target Architecture=FaceNet64, Parameter Ratio (|θc|/|θt|)=34.4/35.42024.05 | 83.41 | 9.33 | 24.33 | 1,909 |