Adversarial Machine Learning for Emotional Privacy

This thesis explores leveraging adversarial machine learning to safeguard emotional privacy in the era of pervasive social media. It proposes a method that introduces subtle modifications, or “adversarial perturbations,” to video-based emotion recognition, making emotions more difficult to detect while preserving natural content. The study examines the applicability of these techniques across multimodal systems combining text, images, and video. A novel architecture for universal adversarial attacks is presented, with evaluations demonstrating its effectiveness in maintaining privacy, robustness, and transferability. This study emphasizes the potential of machine learning in responsibly safeguarding privacy in real-world affective computing applications.