Efficient Model Quality Evaluation in Federated Learning via Functional Encryption

Federated Learning(FL) is a distributed machine learning paradigm that exchanges data among multiple parties without directly sharing the original data. However, FL faces the inherent issue of statistical heterogeneity. Recently, some privacy preserving federated learning scheme have considered this issue. But they use homomorphic encryption, which imposes significant computational and communication overhead on the clients. To address this issue, we propose an efficient model quality evaluation scheme in FL via functional encryption. Specifically, we first use inner product functional encryption(IPFE) to efficiently and securely calculate the cosine similarity between global update and each local update on the server, then use clustering algorithm to assign weights to each client based on cosine similarity, and finally update the global model through weighted aggregation. Experimental results show that compared with other model quality evaluation scheme, our approach increases the computational efficiency by up to 25% and reduces communication cost by up to 70%.