Example: PyTorch + Opacus (DP)

This example trains a small MLP with Differential Privacy using Opacus.

What this is / is not - Is: A real DP-SGD training example with epsilon estimation and accuracy metrics. - Is not: The Aegis orchestration walkthrough. For register/start/status/report, see technical/train_model.md.

When to use - Demonstrate privacy-utility tradeoffs and epsilon accounting with a real model. - Validate DP tuning knobs: noise_multiplier, max_grad_norm, batch_size, epochs.

Prerequisites - Python deps:

pip install torch opacus

Run it

python examples/pytorch_opacus_demo.py --epochs 5 --batch_size 64 --lr 0.1 \
  --noise_multiplier 1.0 --max_grad_norm 1.0 --delta 1e-5 --seed 0

Output (sample)

Train acc: 0.800 | Test acc: 0.770 | Epsilon≈0.XXXX (delta=1e-05)

Notes - Increase noise_multiplier to strengthen privacy (epsilon decreases) but expect more utility loss. - max_grad_norm controls per-sample clipping; tune alongside noise_multiplier. - Need orchestration (participants/strategy/status/report)? See technical/train_model.md.