dp-sgd

Hands-on part of the Federated Learning and Privacy-Preserving ML tutorial given at VISUM 2022

This repo contains my dissertation project done during my Masters in Queen Mary University of London.

Privacy preserving remaining useful life prediction on SCANIA-X truck telemetry: Spectral-DP and DP-SGD under a custom RDP accountant, validated by membership inference attack.

DP-SGD accuracy vs. privacy budget (ε) experiments on MNIST using Opacus — quantifies the accuracy/privacy/compute trade-off with RDP accounting.

A hands-on educational walkthrough of training a CelebA (Eyeglasses) image classifier with Differentially Private SGD using PyTorch and Opacus. The focus of this repo is on clarity and reproducibility through balanced subsets, deterministic preprocessing, and side-by-side baseline vs. DP training, while acknowledging real trade-offs.

This project uses Differentially Private Stochastic Gradient Descent (DP-SGD) and autoencoder-based machine learning to securely analyze smart meter data, protecting sensitive energy patterns while preserving overall trend utility.

Reproducible audit of DP-SGD privacy–utility trade-offs and membership-inference risk for NSL-KDD intrusion detection

Experiments at the intersection of ML security & privacy: adversarial attacks/defenses (FGSM/PGD, adversarial training), differential privacy (DP-SGD, ε–δ), federated learning privacy (secure aggregation), and auditing (membership/model inversion). PyTorch notebooks + eval scripts.

Li X, Chen Y, Wang C, Shen C. When Deep Learning Meets Differential Privacy: Privacy, Security, and More. IEEE Network. 2021 Nov;35(6):148-55.

Rust rewrite of JAX Privacy: DP-SGD primitives, PLD/RDP accounting, matrix factorization, auditing, and adapters.

From-scratch federated learning: FedAvg / FedProx / SCAFFOLD, DP-SGD & secure aggregation, plus FedPer / Byzantine-robust / FedAdam / FedLoRA. 50/50 tests + CI, literature-cross-validated, with honest negative results on Non-IID MNIST.

Membership inference attack plus four layered privacy defenses (DP-SGD, SISA, and more).

Empirical study of the Privacy–Robustness–Performance trilemma in Federated Learning: combining DP-SGD, FLTrust Byzantine-robust aggregation, and Top-k compression across 8 configurations on MNIST, simulated with Flower.

Differentially private LoRA fine-tuning at calibrated epsilon — self-contained RDP accountant + honest, anomaly-aware gating.