Federated learning I

Papers

Sign in to ask or upvote questions. Sign in

1. Full paperPresented

   Asynchronous buffered aggregation for federated learning on heterogeneous edge fleets

   Lucas Meyer (Lakeside Technical University) (presenting), Ana Costa (Lakeside Technical University)

   Abstract

   Synchronous federated rounds stall when slow clients drop out, which is common in vehicle and sensor fleets. We study buffered asynchronous aggregation in which the server updates the global model once a minimum number of client updates has arrived. On an electric-vehicle energy-demand task with 48 clients, the approach reduces wall-clock training time by 41% while keeping mean absolute error within 3% of synchronous training. We report how buffer size and staleness weighting affect convergence and fairness across clients.

   federated learning · asynchronous aggregation · edge devices · energy demand

   Questions from the audience 0

   No questions yet.

2. Full paperPresented

   What does a trusted aggregation server cost? Measuring privacy overheads in federated forecasting

   Oskar Lindqvist (Kestrel Labs) (presenting), Irene Galanis (University of Eastvale), Pablo Moreno (Polytechnic of the North)

   Abstract

   Privacy mechanisms for federated learning are usually compared on accuracy alone. We measure the full cost of three mechanisms on two forecasting tasks: differential privacy, a trusted aggregation server and fully homomorphic encryption with the CKKS scheme. Encrypted updates are about 22 times larger, aggregation time grows by 11 to 15 percent with the trusted server, and differential privacy degrades error most at small client counts. We give guidance on choosing a mechanism from the threat model and the hardware available at the edge.

   federated learning · differential privacy · homomorphic encryption · trusted aggregation

   Questions from the audience 0

   No questions yet.

3. Full paperPresented

   Forgetting on request: certified machine unlearning for federated recommender systems

   Katerina Vlachou (Baltic School of Computing) (presenting), Felix Brandt (Alpine AI Institute), Sara Romano (Northbridge University)

   Abstract

   Data protection law gives users the right to have their data removed, but removing a client's contribution from a federated model is not straightforward. We propose a retraining-free unlearning method for federated matrix factorisation that stores compact per-round summaries and subtracts a client's influence. On two public recommendation datasets the unlearned models match full retraining within 1.2% in recall while using 30 times less computation. We discuss how to verify that forgetting actually happened.

   machine unlearning · federated learning · recommender systems · GDPR

   Questions from the audience 0

   No questions yet.