A Novel Privacy Preserving Framework for Training Dempster-Shafer Theory-Based Evidential Deep Neural Network

Evidential Deep Neural Networks (EDNNs), based on Dempster-Shafer (DS) theory, offer an advanced method for managing uncertainty in deep learning. Like other deep learning models, EDNNs need large datasets, often containing sensitive information, highlighting the need for strong data privacy measures. Our research introduces Pri-DSENN, an innovative framework designed to enhance privacy in the EDNN training process according to DS principles. Pri-DSENN uses a secure multiparty computation (SMC) protocol combined with federated learning (FL) to significantly improve data security. Our experiments with the CIFAR10 and CIFAR100 datasets confirm that integrating SMC with FL in DS-based EDNNs preserves high classification accuracy while effectively handling unclear patterns, ensuring advanced decision-making and robust data privacy and security.