Buffalo: A Practical Secure Aggregation Protocol for Buffered Asynchronous Federated Learning

Federated Learning (FL) has become a crucial framework for collaboratively training Machine Learning (ML) models while ensuring data privacy. Traditional synchronous FL approaches, however, suffer from delays caused by slower clients (called stragglers), which hinder the overall training process. Specifically, in a synchronous setting, model aggregation happens once all the intended clients have submitted their local updates to the server. To address these inefficiencies, Buffered Asynchronous FL (BAsyncFL) was introduced, allowing clients to update the global model as soon as they complete local training. In such a setting, the new global model is obtained once the buffer is full, thus removing synchronization bottlenecks. Despite these advantages, existing Secure Aggregation (SA) techniques—designed to protect client updates from inference

attacks—rely on synchronized rounds, making them unsuitable for asynchronous settings. In this paper, we present Buffalo, the first practical SA protocol tailored for BAsyncFL. Buffalo leverages lattice-based encryption to handle scalability challenges in large ML models and introduces a new role, the assistant, to support the server in securely aggregating client updates. To protect against an actively corrupted server, we enable clients to verify that their local updates have been correctly integrated into the global model. Our comprehensive evaluation—incorporating theoretical analysis and

real-world experiments on benchmark datasets—demonstrates that Buffalo is an efficient and scalable privacy-preserving solution in BAsyncFL environments.