基于同态加密和模型水印的安全可信联邦学习

Secure and trusted federation learning based on homomorphic encryption and model watermarking

为防止联邦学习客户端共享的中间参数泄露,同时保证服务器与客户端的可信性,提出一种结合同态加密和模型水印的联邦学习框架。将Paillier加密技术运用到模型参数的安全聚合中,对参数聚合时的加法同态性进行证明,为提高加密效率在加密前将模型参数进行量化处理;将模型水印技术拓展到安全联邦学习中,利用投影矩阵和正则化函数构建模型水印,将水印模型进行聚合。在MNIST和CIFAR10数据集上的实验验证了提出方法的有效性,提高模型参数加密效率,保证模型的版权。

To prevent the leakage of intermediate parameters shared by federated learning clients and to ensure the trustworthiness between the server and the client,a federated learning framework combining homomorphic encryption and model watermarking was proposed.The Paillier encryption was applied to secure aggregation of model parameters and the additive homomorphism 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.