Explanation Hardening Evaluation on CIFAR-10 10% utility loss, ϵ = 8, δ = 10⁻⁵ (test)
45.44MLS (Pre-Hardening, TPR@0.1% FPR)Anchors
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| AnchorsHardening Method=DeepLeak2026.01 | 45.44 | 14.16 | 75.02 | |
| AnchorsHardening Method=DP-SGD2026.01 | 45.44 | 41.79 | 71.27 | |
| DeconvolutionHardening Method=DeepLeak2026.01 | 16.08 | 0.18 | 30.93 | |
| DeconvolutionHardening Method=DP-SGD2026.01 | 16.08 | 0.52 | 29.38 | |
| ProtoDashHardening Method=DeepLeak2026.01 | 14.9 | 0 | 54.78 | |
| ProtoDashHardening Method=DP-SGD2026.01 | 14.9 | 13.84 | 52.04 | |
| Guided BackPropHardening Method=DeepLeak2026.01 | 10.94 | 1.34 | 42.61 | |
| Guided BackPropHardening Method=DP-SGD2026.01 | 10.94 | 0.8 | 35 | |
| SmoothGradHardening Method=DeepLeak2026.01 | 10.38 | 0.84 | 22.44 | |
| SmoothGradHardening Method=DP-SGD2026.01 | 10.38 | 0.19 | 21.32 | |
| LIMEHardening Method=DeepLeak2026.01 | 10.26 | 0.12 | 3.72 | |
| LIMEHardening Method=DP-SGD2026.01 | 10.26 | 0 | 3.53 | |
| VarGradHardening Method=DeepLeak2026.01 | 6.78 | 0.27 | 6.76 | |
| VarGradHardening Method=DP-SGD2026.01 | 6.78 | 0.14 | 6.42 | |
| GradCAM++Hardening Method=DeepLeak2026.01 | 6.7 | 1.44 | 0.0733 | |
| GradCAM++Hardening Method=DP-SGD2026.01 | 6.7 | 0.8 | 86 | |
| Saliency MapHardening Method=DeepLeak2026.01 | 5.24 | 0.06 | 51.06 | |
| Saliency MapHardening Method=DP-SGD2026.01 | 5.24 | 0.32 | 48 | |
| SHAPHardening Method=DeepLeak2026.01 | 5.04 | 0 | 0.97 | |
| SHAPHardening Method=DP-SGD2026.01 | 5.04 | 0 | 0.92 | |
| Integrated GradientsHardening Method=DeepLeak2026.01 | 4.88 | 0.34 | 0.35 | |
| Integrated GradientsHardening Method=DP-SGD2026.01 | 4.88 | 0.7 | 10 | |
| InputXGradHardening Method=DeepLeak2026.01 | 2.4 | 0.18 | 29.79 | |
| InputXGradHardening Method=DP-SGD2026.01 | 2.4 | 0.8 | 28.3 | |
| OcclusionHardening Method=DeepLeak2026.01 | 1.1 | 0.14 | 2.22 | |
| OcclusionHardening Method=DP-SGD2026.01 | 1.1 | 0.06 | 2.33 | |
| DeepLIFTHardening Method=DeepLeak2026.01 | 1 | 0 | 0.07 | |
| DeepLIFTHardening Method=DP-SGD2026.01 | 1 | 0.06 | 0.08 | |
| GradCAMHardening Method=DeepLeak2026.01 | 0.95 | 0 | 7.86 | |
| GradCAMHardening Method=DP-SGD2026.01 | 0.95 | 0.3 | 23 |