Paper bounds secure aggregation cost when federated-learning users drop out
Original title: Fundamental Limits of Information-Theoretic Secure Aggregation with User Dropouts in Federated Learning
One adaptive scheme meets the optimal masked-gradient communication rate for every feasible dropout set under T-colluding security.
An arXiv paper studies information-theoretic secure aggregation when users drop out of federated learning. It defines key distribution, masked-gradient upload, and mask elimination, and builds one scheme that changes communication with the actual number of dropouts. Converse bounds cover key randomness, key-distribution communication, and both upload phases. The scheme meets the optimal masked-gradient rate in every feasible case and the mask-elimination rate in almost all regimes.
Unconfirmed: The bounds and optimality claims are the authors' theoretical results; no implementation is described.
- Key rate and key-distribution communication are optimal or within a factor of two.
- Mask-elimination bounds are stated for each realized dropout set, not only the worst case.
- Prior schemes typically fix communication for a prescribed dropout threshold.
From the original article
1 min read at arxiv.orgSecure aggregation enables a server to compute the sum of users' gradients or model updates without revealing individual values. With user dropouts, correlated random keys protect local gradients while allowing recovery of the desired gradient sum. We consider three phases: key distribution, masked-gradient upload, and mask elimination.
By Jinbao Zhu · 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.