Phase 2: Attack Results

Baseline and obfuscated test sets evaluation across datasets.

Dataset Deid_Type Param Top1_Acc Top5_Acc
AT&T nan 0 92.5 100
AT&T Pixelization 4 10 30
AT&T Pixelization 8 5 26.25
AT&T Pixelization 16 2.5 13.75
AT&T Gaussian Blur 15 2.5 18.75
AT&T Gaussian Blur 45 3.75 15
AT&T Gaussian Blur 99 3.75 13.75
MNIST nan 0 98.5 99.5
MNIST Pixelization 4 41.5 83.5
MNIST Pixelization 8 5 55
MNIST Pixelization 16 14 46
MNIST Gaussian Blur 15 14 60
MNIST Gaussian Blur 45 14 55.5
MNIST Gaussian Blur 99 14 55.5

Phase 3: Differential Privacy Defense Results

Evaluating impact of Laplacian noise on metrics and model accuracy across datasets.

Dataset Epsilon MSE SSIM Top1_Acc Top5_Acc
AT&T 0.1 17659.9 0.00642367 2.5 15
AT&T 0.3 16299.3 0.0118642 2.5 13.75
AT&T 0.5 15043.9 0.0185997 3.75 13.75
AT&T 0.7 13890.2 0.0251468 3.75 10
AT&T 1 12381 0.0352866 2.5 15
AT&T 3 6181.78 0.107372 5 20
AT&T 5 3466.77 0.187793 2.5 18.75
MNIST 0.1 29345.6 0.0191937 9.5 48
MNIST 0.3 25829.7 0.0586176 8.5 47.5
MNIST 0.5 22773 0.0966827 9.5 50
MNIST 0.7 20113 0.133289 7.5 50
MNIST 1 16784.4 0.185243 9.5 49
MNIST 3 5709.23 0.437811 10 62
MNIST 5 2507.55 0.560655 53.5 88.5

Visualizations

Combined DP vs NP Metrics Plot (Phase 3)

image.png

Phase 1: Baseline De-Identification Visualizations

image.png

image.png

Phase 2: Attack Visualizations (Original vs Obfuscated Predictions)

image.png

image.png

Phase 3: DP Noise Visualizations

image.png

image.png