Privacy-Preserving Federated Learning with Differential Privacy on Census Data

Ten US states hold census records that they may not pool. This project lets them train one income model together with federated learning (FedAvg and FedProx, in a numpy loop and in real Flower runs) and protects every person's record with differential privacy (DP-SGD in each state, epsilon certified by the Opacus accountant). On 883,984 persons of the ACS PUMS 2018 ACSIncome task, FedAvg reaches test accuracy 0.7861 against 0.7871 for a model on the pooled data, and DP-FedAvg stays at 0.7843-0.7875 for epsilon 0.1-8. LightGBM on the pooled data (0.8211) beats every federated and DP logistic model, and local-only models beat FedAvg on these large silos. Fairness, membership- inference and OpenDP audits and a web application complete the study.

Category: Information Science and Technology

Information Science & Technology Project IST09 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.