Image Classification on CIFAR-10-C Brightness Corruption, Severity Level 5 (test)
83.65Accuracy (CIFAR-10-C Brightness 5)No randomization
Evaluation results
| Method | Links | |
|---|---|---|
| No randomizationPrivacy Budget (ϵ)=0.5, δ=10^-5, Classifier=Non-Linear MLP with ReLU, m=64, PA Integration=false2026.05 | 83.65 | |
| No randomizationPrivacy Budget (ε)=0.5, m=642026.05 | 83.65 | |
| No randomizationPrivacy Budget (ε)=1.0, m=642026.05 | 83.65 | |
| No randomizationPrivacy Budget (ε)=1.5, m=642026.05 | 83.65 | |
| No randomizationPrivacy Budget (ε)=2.0, m=642026.05 | 83.65 | |
| No randomizationPrivacy Budget (ε)=2.5, m=642026.05 | 83.65 | |
| No randomizationPrivacy Budget (ε)=3.0, m=642026.05 | 83.65 | |
| No randomizationPrivacy Budget (ε)=3.5, m=642026.05 | 83.65 | |
| No randomizationPrivacy Budget (ε)=4.0, m=642026.05 | 83.65 | |
| No randomizationPrivacy Budget (ε)=4.5, m=642026.05 | 83.65 | |
| No randomizationPrivacy Budget (ε)=5.0, m=642026.05 | 83.65 | |
| No randomizationPrivacy Budget (ε)=7.5, m=642026.05 | 83.65 | |
| No randomizationPrivacy Budget (ε)=10.0, m=642026.05 | 83.65 | |
| PrivUnit2 + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=10.0, m=642026.05 | 58.4 | |
| PrivUnitG + PAPrivacy Budget (ϵ)=10.0, δ=10^-5, PA Integration=true2026.05 | 56.48 | |
| PrivUnitG + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=10.0, m=642026.05 | 56.48 | |
| PrivUnit2 + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=7.5, m=642026.05 | 49.78 | |
| PrivUnitG + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=7.5, m=642026.05 | 48.6 | |
| PrivUnit2 + PAPrivacy Budget (ϵ)=5.0, δ=10^-5, PA Integration=true2026.05 | 37.54 | |
| PrivUnit2 + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=5.0, m=642026.05 | 37.54 | |
| PrivUnitG + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=5.0, m=642026.05 | 36.88 | |
| PrivUnit2Privacy Budget (ε)=10.0, m=642026.05 | 36.15 | |
| PrivUnit2 + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=4.5, m=642026.05 | 34.53 | |
| PrivUnitG + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=4.5, m=642026.05 | 34.14 | |
| PrivUnitGPrivacy Budget (ϵ)=10.0, δ=10^-5, PA Integration=false2026.05 | 33.75 | |
| PrivUnitGPrivacy Budget (ε)=10.0, m=642026.05 | 33.75 | |
| PrivUnit2 + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=4.0, m=642026.05 | 31.52 | |
| PrivUnitG + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=4.0, m=642026.05 | 31.21 | |
| PrivUnit2Privacy Budget (ε)=7.5, m=642026.05 | 29.33 | |
| PrivUnitGPrivacy Budget (ε)=7.5, m=642026.05 | 28.58 | |
| PrivUnit2 + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=3.5, m=642026.05 | 28.4 | |
| PrivUnitG + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=3.5, m=642026.05 | 28.22 | |
| PrivUnit2 + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=3.0, m=642026.05 | 25.36 | |
| PrivUnitG + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=3.0, m=642026.05 | 25.24 | |
| Laplace (ℓ1) + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=10.0, m=642026.05 | 24.16 | |
| PrivUnit2Privacy Budget (ϵ)=5.0, δ=10^-5, PA Integration=false2026.05 | 22.27 | |
| PrivUnit2Privacy Budget (ε)=5.0, m=642026.05 | 22.27 | |
| PrivUnitG + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=2.5, m=642026.05 | 22.27 | |
| PrivUnit2 + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=2.5, m=642026.05 | 22.23 | |
| PrivUnitGPrivacy Budget (ε)=5.0, m=642026.05 | 21.97 | |
| PrivUnit2Privacy Budget (ε)=4.5, m=642026.05 | 20.92 | |
| PrivUnitGPrivacy Budget (ε)=4.5, m=642026.05 | 20.61 | |
| Laplace (ℓ1) + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=7.5, m=642026.05 | 19.89 | |
| PrivUnit2Privacy Budget (ε)=4.0, m=642026.05 | 19.5 | |
| PrivUnit2 + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=2.0, m=642026.05 | 19.36 | |
| PrivUnitG + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=2.0, m=642026.05 | 19.36 | |
| PrivUnitGPrivacy Budget (ε)=4.0, m=642026.05 | 19.31 | |
| CW + Jacobian-Guided Anisotropic Noise Reshaping (PA) w/o boundingPrivacy Budget (ε)=10.0, m=642026.05 | 18.28 | |
| PrivUnit2Privacy Budget (ε)=3.5, m=642026.05 | 18.2 | |
| PrivUnitGPrivacy Budget (ε)=3.5, m=642026.05 | 17.98 | |
| PrivUnit2Privacy Budget (ε)=3.0, m=642026.05 | 16.86 | |
| PrivUnitG + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=1.5, m=642026.05 | 16.72 | |
| PrivUnitGPrivacy Budget (ε)=3.0, m=642026.05 | 16.65 | |
| PrivUnit2 + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=1.5, m=642026.05 | 16.6 | |
| Laplace (ℓ1) + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=5.0, m=642026.05 | 16.09 | |
| CW + Jacobian-Guided Anisotropic Noise Reshaping (PA) w/o boundingPrivacy Budget (ε)=7.5, m=642026.05 | 15.81 | |
| PrivUnit2Privacy Budget (ε)=2.5, m=642026.05 | 15.6 | |
| PrivUnitGPrivacy Budget (ε)=2.5, m=642026.05 | 15.41 | |
| Laplace (ℓ1) + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=4.5, m=642026.05 | 15.38 | |
| Laplace (ℓ1) + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=4.0, m=642026.05 | 14.67 | |
| PrivUnit2Privacy Budget (ε)=2.0, m=642026.05 | 14.34 | |
| Instance OptimalPrivacy Budget (ϵ)=7.5, δ=10^-5, PA Integration=false2026.05 | 14.33 | |
| PrivUnitG + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=1.0, m=642026.05 | 14.25 | |
| PrivUnitGPrivacy Budget (ε)=2.0, m=642026.05 | 14.15 | |
| PrivUnit2 + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=1.0, m=642026.05 | 14.1 | |
| Laplace (ℓ1) + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=3.5, m=642026.05 | 14.05 | |
| CW + Jacobian-Guided Anisotropic Noise Reshaping (PA) w/o boundingPrivacy Budget (ε)=5.0, m=642026.05 | 13.66 | |
| Laplace (ℓ1) + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=3.0, m=642026.05 | 13.42 | |
| CW + Jacobian-Guided Anisotropic Noise Reshaping (PA) w/o boundingPrivacy Budget (ε)=4.5, m=642026.05 | 13.23 | |
| PrivUnit2Privacy Budget (ε)=1.5, m=642026.05 | 13.16 | |
| PrivUnitGPrivacy Budget (ε)=1.5, m=642026.05 | 12.99 | |
| PLAN: Variance-AwarePrivacy Budget (ϵ)=5.0, δ=10^-5, PA Integration=false2026.05 | 12.9 | |
| CW + Jacobian-Guided Anisotropic Noise Reshaping (PA) w/o boundingPrivacy Budget (ε)=4.0, m=642026.05 | 12.82 | |
| Laplace (ℓ1) + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=2.5, m=642026.05 | 12.8 | |
| Laplace (ℓ1)Privacy Budget (ε)=10.0, m=642026.05 | 12.51 | |
| CW + Jacobian-Guided Anisotropic Noise Reshaping (PA) w/o boundingPrivacy Budget (ε)=3.5, m=642026.05 | 12.41 | |
| Laplace + PAPrivacy Budget (ϵ)=2.0, δ=10^-5, PA Integration=true2026.05 | 12.19 | |
| Laplace (ℓ1) + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=2.0, m=642026.05 | 12.19 | |
| PrivUnit2Privacy Budget (ε)=1.0, m=642026.05 | 12.03 | |
| PrivUnitG + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=0.5, m=642026.05 | 12.01 | |
| CW + Jacobian-Guided Anisotropic Noise Reshaping (PA) w/o boundingPrivacy Budget (ε)=3.0, m=642026.05 | 12.01 | |
| PrivUnit2 + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=0.5, m=642026.05 | 11.94 | |
| PrivUnitGPrivacy Budget (ε)=1.0, m=642026.05 | 11.9 | |
| Laplace (ℓ1)Privacy Budget (ε)=7.5, m=642026.05 | 11.77 | |
| CW + Jacobian-Guided Anisotropic Noise Reshaping (PA) w/o boundingPrivacy Budget (ε)=2.5, m=642026.05 | 11.64 | |
| Laplace (ℓ1) + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=1.5, m=642026.05 | 11.61 | |
| AGM + ℓ2-clippingPrivacy Budget (ϵ)=3.5, δ=10^-5, PA Integration=false2026.05 | 11.57 | |
| CW + Jacobian-Guided Anisotropic Noise Reshaping (PA) w/o boundingPrivacy Budget (ε)=2.0, m=642026.05 | 11.28 | |
| Laplace (ℓ1)Privacy Budget (ε)=5.0, m=642026.05 | 11.11 | |
| Laplace (ℓ1) + Jacobian-Guided Anisotropic Noise Reshaping (PA)Privacy Budget (ε)=1.0, m=642026.05 | 11.03 | |
| Laplace (ℓ1)Privacy Budget (ε)=4.5, m=642026.05 | 10.99 | |
| PrivUnit2Privacy Budget (ε)=0.5, m=642026.05 | 10.99 | |
| CW + Jacobian-Guided Anisotropic Noise Reshaping (PA) w/o boundingPrivacy Budget (ε)=1.5, m=642026.05 | 10.92 | |
| Laplace (ℓ1)Privacy Budget (ε)=4.0, m=642026.05 | 10.87 | |
| PrivUnitGPrivacy Budget (ε)=0.5, m=642026.05 | 10.83 | |
| Laplace (ℓ1)Privacy Budget (ε)=3.5, m=642026.05 | 10.74 | |
| Laplace (ℓ1)Privacy Budget (ε)=3.0, m=642026.05 | 10.63 | |
| CW + PA w/o boundingPrivacy Budget (ϵ)=1.0, δ=10^-5, PA Integration=true2026.05 | 10.58 | |
| CW + Jacobian-Guided Anisotropic Noise Reshaping (PA) w/o boundingPrivacy Budget (ε)=1.0, m=642026.05 | 10.58 | |
| Laplace (ℓ1)Privacy Budget (ε)=2.5, m=642026.05 | 10.5 |