Towards Privacy-Preserving Split Learning: Destabilizing Adversarial Inference and Reconstruction Attacks in the Cloud

Abstract

This work aims to provide both privacy and utility within a split learning framework while consideringboth forward attribute inference and backward reconstruction attacks. To address this, a novelapproach has been proposed, which makes use of class activation maps and autoencoders as a plug-instrategy aiming to increase the user’s privacy and destabilize an adversary. The proposed approachis compared with a dimensionality-reduction-based plug-in strategy, which makes use of principalcomponent analysis to transform the feature map onto a lower-dimensional feature space. Our workshows that our proposed autoencoder-based approach is preferred as it can provide protection at anearlier split position over the tested architectures in our setting, and, hence, better utility for resource constraineddevices in edge-cloud collaborative inference (EC) systems.

Keywords: Split Learning, Edge-cloud Collaborative Systems, Privacy-Preserving Learning, Autoencoder, Dimensionality Reduction, Privacy and Utility