Privacy-Preserving High-Dimensional Distributed Learning Via Client-Side Deep Denoising Sparse Autoencoders, Adaptive Attribute Filtering, And Federated Split Learning With Gradient Compression

Abstract

With high-dimensional data sets, especially, privacy preservation (PP) in deep learning (DL) is gaining increasing importance due to the significant computational challenges. Typical methods often have the 'curse of dimensionality', leading to inefficiency and privacy risks. This extended study makes two novel algorithmic contribution: (1) Adaptive Data Attribute Filtering (ADAF): replace the entropy-based filtering with a combined entropy-mutual-information scoring and redundancy-pruning strategy and (2) Federated DDSAE with Adaptive Gradient Compression (FDAG): top-k gradient compression for communication efficiency during multi-client parallel training. Experiments on the following datasets showcased the following improvements of the extended CDDSAE-EXT framework over the original CDDSAE baseline: On the PTB-XL ECG dataset, the extended CDDSAE-EXT framework improves the latency by 22.4%, the running time by 46.1%, and the minimum information loss by 0.17, across three independent comparison tables.