To prevent the leakage of intermediate parameters shared by federated learning clients and to ensure the trustworthi-ness between the server and the client,a federated learning framework combining homomorphic encryption and model water-marking was proposed.The Paillier encryption was applied to secure aggregation of model parameters and the additive homomor-phism in parameter aggregation was proved,while model parameters were quantified before encryption to improve encryption efficiency.The model watermarking technique was extended to secure federal learning by constructing model watermarks using projection matrices and regularization functions and aggregating the watermarked models.Experiments on the MNIST and CIFAR10 datasets validate the effectiveness of the proposed method,the encryption efficiency of model parameters is improved and the copyright of the models is ensured.
关键词
联邦学习/安全可信/参数量化/模型聚合/同态加密/投影矩阵/模型水印
Key words
federated learning/secure and trusted/parameter quantification/model aggregation/homomorphic encryption/pro-jection matrix/model watermarking
维普
万方数据