FedSumUp shares privacy-safe compressed data for federated learning
FedSumUp: User-Auditable Privacy-Aware Representation Sharing for Resource-Constrained Federated Intelligence
Cryptography and SecurityArtificial Intelligence
Summary
Sharing data from many devices to improve AI can risk revealing private information. The authors propose FedSumUp, a method that compresses data into simpler, less revealing forms before sending it to a central server. This compression balances privacy and usefulness, letting users check how much detail is shared. Their approach reduces the amount of work each device must do and still keeps AI models accurate, even when data varies across devices.
What this means in practice
- •For mobile device developers: Enable privacy-aware on-device data compression to reduce processing time and bandwidth in federated learning systems.
- •For healthcare it teams: Deploy privacy-preserving data sharing methods for AI training on sensitive medical images without exposing patient details.
Authors
Jingwei Zhao, Tiantong Sun, Yuanyuan Song, Jinhuan Guo, Wei Zishuo, Shengyu Huang, Ziyuan Ren, Wei Chen, Xiaojin Zhang
Abstract
Existing 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. To address these issues, we propose FedSumUp, which employs a learnable lossy compressor on clients to generate compact and visually interpretable representations. By controlling information capacity, FedSumUp reduces privacy leakage while allowing users to directly assess the privacy-utility tradeoff. We further introduce a dual-objective optimization mechanism to preserve task-relevant semantics while suppressing reconstruction risks. The compressor is trained on the server and only performs lightweight inference on clients. Since the server cannot access raw data, we develop a data-free closed-loop optimization strategy that optimizes the compressor by minimizing the discrepancy between original and twice-compressed representations after structure-preserving upsampling. Experiments on BloodMNIST, CIFAR-10, and ChestMNIST demonstrate reconstruction PSNR as low as 8.56 dB (with synthetic PSNR as low as 5.45 dB), client runtime reductions up to 96.5% (from 778s to 27s on ChestMNIST), and competitive classification accuracy (89.72% ChestMNIST, 82.24% BloodMNIST, 65.63% CIFAR-10), while remaining inherently robust to non-IID data distributions.