Emotion detection with privacy preservation using adversarial learning
Abstract
The continuous monitoring of one's emotional state can provide valuable insights about their psychological well-being and can be used as a foundation for diagnosis and treatment applications. Yet, due to privacy concerns, technologies that continuously monitor signals that reflect emotions, such as images, are met with strong skepticism. This thesis aims to design a privacy-preserving image generation algorithm that anonymizes the input image and at the same time maintains emotion-related information. To do so, we identify landmarks in human faces and quantify the amount of emotion and identity based information carried by each of the landmarks. We then propose a modification of a conditional generative adversarial network that can transform facial images in such a way that the identity based information is ignored while the emotion based information is retained. We then evaluate the degree of emotion and identity content in the transformed images by performing emotion and identity classification using these images. The proposed system is trained and evaluated on two publicly available datasets, namely the Yale Face Database and the Japanese Female Facial Expression dataset, and the generated images achieve moderate to high emotion classification accuracy and low identity classification accuracy.
Citation
Ramesh, Ravikiran (2021). Emotion detection with privacy preservation using adversarial learning. Master's thesis, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /196090.