K-Anonymization and Residual Neuron Attention Network for Privacy Data Protection in Blockchain Network With Federated Learning Using Defense Application

Abstract

Blockchain acts as an important potential in the defense applications for several defense uses because of its features, namely transparency, decentralization, immutability, and security. However, protecting privacy data has various security liabilities and attack issues. Therefore, a new model named Residual Neuron Attention Network (ResNA-Net) has been devised for privacy data protection in defense applications. In Federated Learning (FL), the entities, like server and nodes are included. Here, local training is done and the weights are updated to the server first, and next, model aggregation at the server is executed. Then, the global model is downloaded at all nodes, training is updated, and the process is iterated at all epochs. Meanwhile, in local training, the input defense data is normalized by Min-max normalization and then augmented using oversampling. Then, k-anonymization is executed using Fractional Gradient Beluga Whale Optimization (FGBWO). Next, privacy-protected data classification is executed by employing ResNA-Net, which is engineered by the combination of Deep Residual Network (DRN) and Neuron Attention Stage-by-Stage Net (NasNet). The ResNA-Net achieved high performance and the immutable nature of the blockchain used in the ResNA-Net model protects the defense data during the entire process of the system. The hybrid ResNA-Net effectively learns the complex features and this capability improves the accuracy of the model. The high-performance results obtained by the devised model highly protect sensitive data thereby providing security and privacy in defense data applications.