In this work we propose a data-driven adversarial learning framework for generating compressing representations that retain utility comparable to state-of the-art. This is achieved by applying adversarial learning scheme to the design of compression network, whose utility/privacy performances are evaluated by the utility classifier and the adversary reconstructor, respectively.

We demonstrate CPGAN achieves the best utility/privacy trade-off on the benchmark dataset in comparison with the previous work. We also demonstrate that CPGAN attains comparable utility accuracy whilst resisting the reconstruction attack on the real image dataset assuming white-box attack. We incorporate the funnel layer into CPGAN, thus enabling the compression of raw data into lower dimension.

This technology aims to establish a privacy preserving mechanism for data acquisition during the machine learning process to encourage enterprises to provide information to the AI Innovation Center in a safe and reliable manner, so that the AI Innovation Center can be obtained in the past with difficulty, personal or corporate privacy information to promote industry 4.0 and smart manufacturing related research.