Learning Privacy-Preserving Embeddings for Image Data to Be Published

Deep learning has its superiority on learning feature representations that bring promising performance in various application domains. Recent advances have shown that privacy attributes of users and patients (e.g., identity, gender and race) can be accurately inferred from image data. To avoid the risk of privacy leaking, data owners can resort to release the embeddings, rather than the original images. In this paper, we aim at learning to generate privacy-preserving embeddings from image data. The obtained embeddings are required to maintain the data utility (e.g., keep the performance of the main task like disease prediction), and to simultaneously prevent the private attributes of data instances from being accurately inferred. In addition, we also want the embeddings hard to be successfully used to reconstruct the original images. We propose a hybrid method based on multi-task learning to reach the goal. The key idea is two-fold. One is to learn the feature encoder that can benefit the main task and fool the sensitive task at the same time via iterative training and feature disentanglement. The other is to incorporate the learning of adversarial examples to mislead the sensitive attribute classification’s performance. Experiments conducted on Multi-Attribute Facial Landmark (MAFL) and NIH Chest X-rays datasets exhibit the effectiveness of our hybrid method. A set of advanced studies also show the usefulness of each model component, the difficulty in data reconstruction, and the performance impact of task correlation.