Secure aggregation (SA) has emerged as a vital component of federated learning (FL), enabling collaborative training of a global machine learning model while safeguarding the privacy of clients’ local datasets. Most existing SA protocols implement the privacy preserving-variant of federated averaging (FedAvg) as the aggregation technique and assume independent and identically distributed (IID) datasets across clients. This assumption makes FedAvg unsuitable for non-IID scenarios, where variations in client datasets lead to less effective global model. We propose SAAFL, a SA protocol specifically designed for non-IID settings and more specifically for the recently proposed federated label-aware aggregation (FedLA) protocol. SAAFL computes the weighted average of clients’ inputs where weights depend on the label distributions and should remain confidential. SAAFL is resilient to client dropouts and supports client selection. Our experimental results show that it achieves comparable model accuracy with FedLA and remains efficient in terms of computation and communication.