Various automated services may be improved by understanding the psychological state their users are in. Spotify, a music streaming service, for example, could suggest music matching the listeners’ mood. Smart-devices are increasingly being installed in homes and their services could benefit from the emotion recognition of the occupants. However, this benefit may require data to be transmitted for processing externally. This creates a concern around privacy-preservation. Consumers are reluctant to purchase devices in which personal information is exposed, particularly in the form of visual or audio data. They worry that private conversations may be stored and shared without their explicit permission.
In this project, we extended prior work on privacy-preservation using adversarial training of deep artificial neural networks (Wu, Wang, 2018) to the domain of emotion recognition from audio and visual inputs. We showed that subject identity attributes can be largely eliminated while maintaining emotion classification performance.
Z. Wu, Z. Wang, Z. Wang, and H. Jin. Towards privacy-preserving visual recognition via adversarial training: A pilot study. In Proceedings of the European Conference on Computer Vision (ECCV), pages 606–624, 2018.
IEMOCAP dataset for the project can be retrieved from: https://sail.usc.edu/iemocap/