Straggler-Resilient Secure Aggregation for Federated Learning

We present CodedSecAgg, a straggler-resilient secure aggregation scheme for federated learning. CodedSecAgg intro-duces redundancy on the devices' data across the network, which is leveraged during the iterative learning phase at the central server to update the global model based on the responses of a subset of the devices. Compared to other schemes in the literature, which deal with device dropouts by ignoring the contribution of dropped devices, the proposed scheme does not suffer from the client-drift problem. We apply CodedSecAgg to a classification problem on the MNIST dataset. For a scenario with 120 devices, we show that CodedSecAgg outperforms state-of-the-art LightSecAgg in terms of latency by a factor of 6.6 to 15.8, depending on the number of colluding agents, for an accuracy of 95%.