FedSumUp: User-Auditable Privacy-Aware Representation Sharing for Resource-Constrained Federated Intelligence
FedSumUp shares compact client representations to cut federated privacy leakage and client compute.
FedSumUp is a federated representation-sharing method that uses a learnable lossy compressor so resource-constrained clients can share compact, visually interpretable features instead of synthetic data. A dual-objective optimizer keeps task semantics while lowering reconstruction risk, and a data-free closed-loop trains the compressor without server access to raw data. On BloodMNIST, CIFAR-10, and ChestMNIST, reported reconstruction PSNR is as low as 8.56 dB, client runtime drops up to 96.5%, and classification accuracy reaches 89.72%, 82.24%, and 65.63% respectively.
- Clients upload compact lossy representations instead of synthetic data.
- Server-trained compressor; clients only run lightweight inference.
- ChestMNIST client time fell from 778s to 27s, a 96.5% cut.
- Reconstruction PSNR as low as 8.56 dB; accuracy stays competitive.
From the original article
1 min read at arxiv.orgExisting federated data condensation methods face privacy risks, as uploaded synthetic data may preserve sensitive information and remain vulnerable to reconstruction attacks. Moreover, their generation often incurs considerable client-side computational overhead, limiting deployment on resource-constrained devices.
By Jingwei Zhao · 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.