DP-SGD Interactive Playground

Model Configuration

Dataset

Model Architecture

Quick Presets

DP-SGD Parameters

Clipping Norm (C)

0.1 1.0 5.0

Noise Multiplier (σ)

0.1 1.0 5.0

Batch Size

16 64 512

Learning Rate (η)

0.001 0.01 0.1

Epochs

1 30 50

Estimated Privacy Budget (ε) ? This is the estimated privacy loss from training with these parameters. Lower ε means stronger privacy guarantees.

2.47

Stronger Privacy Weaker Privacy

Training Progress

Training Metrics

Gradient Clipping

Privacy Budget

View:

Showing 5 data points

Gradient Clipping Visualization

The chart below shows a distribution of gradient norms before and after clipping. The vertical red line indicates the clipping threshold. ? Clipping ensures no single example has too much influence on model updates, which is essential for differential privacy.

Privacy Budget Consumption

This chart shows how the privacy budget (ε) accumulates during training. ? In differential privacy, we track the 'privacy budget' (ε) which represents the amount of privacy loss. Lower values mean stronger privacy guarantees.

Results

Final Metrics

Recommendations

Run training to see results here

📊

92.4%

Accuracy

0.283

Loss

ε = 2.1

Privacy Budget

3.7s

Training Time

Privacy-Utility Trade-off

Utility

Privacy

This model achieved 92.4% accuracy with a privacy budget of ε=2.1. This is a good trade-off for most applications.