A Privacy-Preserving Multi-Task Learning Framework For Emotion and Identity Recognition from Multimodal Physiological Signals

The increasing popularity of empathetic sensors can play a significant role in the affective computing era. Recognizing human emotion from these unobtrusive devices is an important building block in this context. Multi-task learning has been studied extensively for various machine learning tasks, including affective computing, which uses the shared information between related tasks to improve performance. Since the physiological data from the mentioned sensors contain private data, they can also lead to privacy threats by exposing highly sensitive information. To address this issue, we combine differential privacy and federated learning approaches with multi-task learning to efficiently recognize the user's mental stress while perturbing private user identity information. More concretely, the proposed framework improves the performance of emotion recognition tasks by taking advantage of multi-task learning and preserving privacy. We extensively evaluate our framework with the employed dataset: results show an accurate emotion recognition of 90% while limiting the re-identification accuracies to 47%.