DealSecAgg: Efficient Dealer-Assisted Secure Aggregation for Federated Learning

Federated learning eliminates the necessity of transferring private training data and instead relies on the aggregation of model updates. Several publications on privacy attacks show how these individual model updates are vulnerable to the extraction of sensitive information. State-of-the-art secure aggregation protocols provide privacy for participating clients, yet, they are restrained by high computation and communication overhead. We propose the efficient secure aggregation protocol DealSecAgg. The cryptographic scheme is based on a lightweight single-masking approach and allows the aggregation of the global model under encryption. DealSecAgg utilizes at least one additional dealer party to outsource the aggregation of masks and to reduce the computational complexity for mobile clients. At the same time, our protocol is scalable and resilient against client dropouts. We provide a security proof and experimental results regarding the performance of DealSecAgg. The experimental evidence on the CIFAR-10 data set confirms that using our protocol, model utility remains unchanged compared to plain federated learning. Furthermore, the results show how our work outperforms other state-of-the-art masking strategies both in the number of communication rounds per training step and in computational costs, which grows linearly in the amount of active clients. By employing our protocol, runtimes can be reduced by up to 87.8% compared to related work.