Federated data condensation is a technique that lets clients share compact synthetic data instead of raw gradients, but it carries privacy risks. According to the arXiv abstract, uploaded synthetic data may preserve sensitive information and remain vulnerable to reconstruction attacks. The abstract also notes that generating this condensed data often incurs considerable computational cost, which is a barrier for resource-constrained participants.
FedSumUp is proposed as a response to these problems. The paper describes a "user-auditable privacy-aware representation sharing" method, suggesting that it gives users a way to inspect or verify how their data is shared while keeping the representation private. The approach is explicitly designed for resource-constrained federated intelligence, so it likely addresses both the privacy and efficiency shortcomings of existing methods.
The abstract is truncated, so the full technical details of FedSumUp are not available from this source alone. Still, the stated goals are clear: mitigate privacy leakage, add auditability, and lower the cost of generating condensed data. As a preprint, the work has not yet undergone peer review, but it points to an active area of research in privacy-preserving federated learning.