Efficient Mobile-Cloud Collaborative Aggregation for Federated Learning with Latency Resilience

With the rapid growth of mobile and edge computing, federated learning (FL) has emerged as a key technology to enable collaborative model training on mobile devices while preserving user privacy. Secure aggregation is an essential component in FL to protect local gradients and compute the global model, but it is vulnerable to threats from high-latency. When some clients arrive late, the pairwise masks among clients cannot be canceled properly, forcing the server to learn the late clients' masks in order to complete the aggregation. As a result, network uncertainty puts the aggregation process at risk of either service interruption or privacy leakage. While double masking is treated as the most effective solution to achieve both robustness and privacy, its computational and communication costs are prohibitive, especially for resource-constrained mobile devices. To address these challenges, we propose an Efficient Mobile-Cloud Collaborative Aggregation for Federated Learning with Latency Resilience (EFL-LR). We leverage Shamir's secret sharing and a key-homomorphic pseudorandom function to ensure privacy for high-latency clients while reducing computation overheads to O(n\log ² n + d) for clients and O(n+d) for the server. Formal security analysis confirms its latency resilience and privacy guarantees. Experimental results show that EFL-LR achieves 2- 3× lower client-side computation cost and accelerates server-side aggregation recovery by at least 10×.