Efficient federated learning privacy protection scheme

Federated learning allows clients to jointly train models with only shared gradients, rather than directly feeding the training data to the server. Although federated learning avoids exposing data directly to third parties and plays a certain role in protecting data, research shows that the transmission gradient in federated learning scenarios will still lead to the disclosure of private information. However, the computing and communication overhead brought by the encryption scheme in the training process will affect the training efficiency, and it is difficult to apply to resource-constrained environments. Aiming at the security and efficiency problems of privacy protection schemes in current federated learning, a safe and efficient privacy protection scheme for federated learning is proposed by combining homomorphic encryption and compression techniques. The homomorphic encryption algorithm is optimized to ensure the security of the scheme, reduce the number of operations and improve the efficiency of operations. At the same time, a gradient filtering compression algorithm is designed to filter out the local updates that are not related to the convergence trend of the global model, and the update parameters are quantized by a computationally negligible compression operator, which ensures the accuracy of the model and increases the communication efficiency. The security analysis shows that the scheme satisfies the security characteristics such as indistinguishability, data privacy and model security. Experimental results show that the proposed scheme has not only higher model accuracy, but also obvious advantages over the existing schemes in terms of communication cost and calculation cost. © 2023 Science Press. All rights reserved.