Assess students' ability to apply differential privacy methods—especially DP-SGD—to real ML training workflows: configuring clipping norms, noise multipliers, batch sizes and epochs; computing privacy loss (ε, δ) via RDP/moments accountant and amplification by subsampling; implementing per-example gradients and microbatching; evaluating utility–privacy tradeoffs and model performance; handling non‑IID/federated data and deployment/debugging considerations to meet specified privacy budgets and accuracy targets.