Abstract
Federated learning (FL) keeps training data local, but it does not by itself prevent attacks against the shared model. Reconstruction and inference attacks can still recover information from model parameters and learned representations, which is a serious concern for medical imaging. We study Gerchberg-Saxton (GS) spectral preprocessing, a stochastic phase-randomization and frequency-masking transform applied on each client before training. We frame GS strictly as an empirical leakage-mitigation step and make no formal privacy claims. Using a simulated three-client FL setup on brain MRI under IID and non-IID partitions, we evaluate GS with three attacks: a blind feature-inversion attack with a matched-versus-wrong separability protocol; a query-only black-box semantic-leakage attack; and a membership-inference check. DP-SGD is included only as a formal-privacy reference point. Baseline models leak measurable per-sample structure and recoverable class signal. GS reduces both toward chance while preserving classification utility at low and moderate masking. The most protective masking level is attack-dependent, and the response is non-monotonic, so we report effects without claiming a dose-response relationship. We also implement GS as a client-side package in a standard FL framework to show deployment feasibility. GS does not provide an (ε,δ) guarantee and is presented as an empirical, low-cost preprocessing step rather than a formal privacy mechanism.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DeCaF_002.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/DeCaF_002_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=UB5ptOfJ0H
BibTex
@InProceedings{AySeh_GerchbergSaxton_MICCAISAT2026,
author = { Ay, Seha AND Zhang, Wei AND Topaloglu, Umit},
title = { { Gerchberg-Saxton Spectral Preprocessing as Empirical Leakage Mitigation in Simulated Federated Medical Imaging } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17252},
month = {pending},
page = {pending}
}