Research on a Hybrid Security Model for Distributed Neural Networks

This research proposes a hybrid method for protecting distributed Machine Learning (ML) systems that combines neural network interconnection strengthening with model architecture obfuscation. Experimental results demonstrate that the combined approach maintains the accuracy of the global learning task while outperforming isolated methods (obfuscation-only or strengthening-only). The proposed method significantly enhances the robustness of inter-network communication by preserving the consistency of feature transmission among neural networks. Both obfuscation and hybrid strategies significantly reduce the effectiveness of various model attacks, such as model stealing, model inversion, and membership inference, thereby lowering the fidelity score of surrogate models. The overall evaluation indicates that the proposed approach achieves a well-balanced trade-off between accuracy, confidentiality, and attack resistance within the defined system constraints.