Federated Learning-Based Privacy Protection Scheme for Intelligent Medical Assessment

With the development of the Healthy China and Data China strategy, healthcare big data has become a national core asset, but the data faces the risk of privacy leakage. Federated learning can obtain more accurate training models without sharing data, which can reduce the risk of local data privacy leakage, but it cannot resist the current latest model attack methods for federated learning framework, such as model reverse, etc.For this reason, this paper proposes a privacy protection scheme for intelligent medical assessment based on federated learning. To address the privacy leakage problem faced by medical data during federated learning, a federated learning architecture based on the LIP adaptive differential privacy algorithm is proposed to protect the privacy security of the whole process of health data during federated learning. Secondly, to address the difficulties of accuracy reduction and high computational overhead of edge devices caused by differential privacy, the lightweight model MobileNet-V2 is selected as the backbone network, and the LSTM mechanism is designed to maintain the state information of the features to enhance the performance of the model. The scheme achieves 96.3 % accuracy on the Maternal Health Risk Data dataset and reduces the original matching rate to 22%, which ensures higher model accuracy while effectively resisting gradient leakage attacks.