Adaptive Dynamic Adjustment of Privacy Protection Intensity Under Federated Learning

As a distributed machine learning method, federated learning can solve the conflict between data privacy and data sharing in machine learning. It can also be combined with techniques such as homomorphic encryption to further improve data security. However, there are still two problems in the actual situation. Firstly, there is still a risk of transmission and leakage of data such as model parameters and embedding matrices in the federated learning process. Secondly, the degree of familiarity and trust between participants in federated learning are different, and different data require different levels of privacy security protection. If complex privacy protection techniques are used in scenarios with low security protection requirements, it will lead to unnecessary performance degradation of the algorithm. To solve these problems, we first propose a four-level privacy protection intensity classification strategy combining federated learning and homomorphic encryption technology. Then we study the mechanism and principle of homomorphic encryption technology combined with horizontal and vertical federated learning respectively, and explore two federated learning frameworks that adaptively and dynamically adjust the privacy protection intensity. It includes the federated homomorphic encryption model weight matrix framework and the federated homomorphic encryption embedding representation framework, so as to realize privacy protection at both the model level and the data level. Finally, the effectiveness of the framework is verified by experiments.