Direction-Shaped Anisotropic Pre-DP Perturbation for Privacy-Preserving Federated Medical Imaging
Custom code for the manuscript "Direction-shaped anisotropic pre-DP gradient perturbation and clinically grounded gradient-inversion evaluation for differentially private federated medical imaging." Implements a three-tier federated learning pipeline (FedProx + Opacus DP-SGD + a pre-DP perturbation shield) and evaluates whether shaping the geometry of the pre-DP noise, using an anisotropic correlated covariance built from the two-body correlators of a four-qubit entangled circuit, changes utility, privacy budget, or gradient-inversion resistance relative to isotropic noise and a variance-matched classical-correlated control. Includes membership-inference and gradient-inversion (iDLG and Geiping) attack evaluation, Rényi-DP accounting, and an optional 128x128 clinical-resolution validation phase. Datasets (PathMNIST, OrganAMNIST) are downloaded from the public MedMNIST benchmark; no patient data are included.