SDV-Aggregation: Client Dual Validation for Secure Aggregation

Secure aggregation enables servers to learn the sum of client vectors in a privacy-preserving manner, which has been widely used in federated learning. To date, existing schemes primarily focus on ensuring security during the aggregation process but fall short in effectively validating client inputs and computations. This limitation hampers the server's ability to accurately assess the validity of client contributions, ultimately impacting the performance of machine learning models. To address these challenges, we introduce SDV-Aggregation, a novel client dual-validation framework for secure aggregation. For input validation, we propose a data validation mechanism of vector commitment based on the message authentication code to verify the validity of input data. In order to ensure the correlation of model input and output, We propose a new remote validation method for local training based on Trusted Execution Environment (TEE) to verify the correctness of the model training process and results. Experimental results show that, compared to state-of-the-art secure aggregation schemes, SDV-Aggregation achieves superior security performance while maintaining minimal communication and computation overhead. Moreover, it demonstrates strong adaptability in FL environments with non-independent and identically distributed (Non-IID) data for the client.