| Summary: | Neural Network, especially its variant, Convolution Neural Network, demonstrated huge potential in terms of processing image data. However, the spread of AI technology has raised the public’s concern about technology misuse, including the risk of privacy leakages. In-vehicle monitoring is such a scenario. On the one hand we expect a monitoring system to detect abnormal actions within the vehicle, while on the other hand, we do not want to give up our privacy. In this thesis, we first review recent developments of ML technology, and then introduce our target application scenario: in-vehicle monitoring. Next, we review the existing privacy preserving technology and we eventually proposed an approach that can be applied in our target scenario. This privacy-preserving framework can learn from a target machine learning task and generate a sifted data representation which only contains essential features for that specific task. We prove that this framework can provide considerable privacy protection with an acceptable accuracy loss of 6.34%. However, further experiments might be need to evaluate performance of the framework within in-vehicle scenario, as Cloak suffers higher accuracy loss of 13.13% in this scenario.
|
|---|