PPFLex uses MPC to stop federated guidance from leaking labels
Original title: Privacy-Preserving Federated Loss Exploration for Vision-Based Perception under Non-IID Data
In the authors' experiment, a malicious server inferred client label sets from shared guidance matrices at 50% to 84% accuracy.
The paper says gradient-guided federated learning for vehicle cameras improves non-IID training but exposes a privacy weakness when guidance matrices are shared. Their leakage experiment used Dirichlet label partitions and a malicious server that inferred clients' label sets. PPFLex replaces the central server with three MPC servers that compute the global guidance matrix and federated averaging under semi-honest and malicious assumptions. On GTSRB and BTSC, they report accuracy comparable to plaintext FedLEx with acceptable MPC overhead.
Unconfirmed: Leakage rates come from the authors' simulated partitions, not a confirmed production incident.
- Label inference accuracy in the experiment ranged from 50% to 84%.
- PPFLex uses three MPC servers rather than one central server.
- The threat model covers semi-honest and malicious MPC servers.
- Reported benchmarks are the GTSRB and BTSC traffic-sign datasets.
From the original article
1 min read at arxiv.orgVision-based image classification models benefit from Federated Learning (FL) across vehicles, however learning accurate models from distributed vehicular cameras is challenging due to data heterogeneity and privacy concerns. To improve performance on non-IID data, advanced FL optimization approaches have emerged that share auxiliary insights, such as client gradient behaviors.
By Nergiz Yuca · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.