Feasibility of reconstruction attacks on Deep Neural Networks when parts of the model are exposed to attackers

Identifier:  TFM:2418

Handle: https://hdl.handle.net/20.500.11797/TFM2418

Authors:  Prieto Tárrega, Hugo

Abstract:
This work analyzes reconstruction attacks against partially protected deep neural networks (DNNs) in Trusted Execution Environments (TEEs), where some layers are exposed. Using ResNet-50 trained by EMBL as a case study, black-box (generative inversion) and white-box (layer-by-layer inversion) attacks are implemented to evaluate the recovery of original images. Visual similarity, training time, and classification accuracy are measured. The results reveal privacy risks associated with partial protections and aim to guide more robust techniques in deep learning.