OpticalDR: A Deep Optical Imaging Model for Privacy-Protective Depression Recognition

TL;DR

This work tackles privacy concerns in automatic depression recognition by proposing OpticalDR, a hardware-software system that erases identity information at capture while preserving depression-relevant cues. It combines a learnable thin-lens optical module with a SANet-based emotion and depression feature extractor, trained end-to-end in a progressive, multi-task framework. The optical lens is optimized jointly with downstream models to prevent facial identity leakage while maintaining predictive power for depression, and evaluation on CelebA, AVEC 2013, and AVEC 2014 shows strong privacy protection (average face-recognition AUC around 0.51) and competitive depression scores (MAE/RMSE in the mid-to-high sevens). The approach offers robust privacy guarantees, because no identifiable facial images are stored or transmitted, with potential for physical deployment and multi-modal extensions.

Abstract

Depression Recognition (DR) poses a considerable challenge, especially in the context of the growing concerns surrounding privacy. Traditional automatic diagnosis of DR technology necessitates the use of facial images, undoubtedly expose the patient identity features and poses privacy risks. In order to mitigate the potential risks associated with the inappropriate disclosure of patient facial images, we design a new imaging system to erase the identity information of captured facial images while retain disease-relevant features. It is irreversible for identity information recovery while preserving essential disease-related characteristics necessary for accurate DR. More specifically, we try to record a de-identified facial image (erasing the identifiable features as much as possible) by a learnable lens, which is optimized in conjunction with the following DR task as well as a range of face analysis related auxiliary tasks in an end-to-end manner. These aforementioned strategies form our final Optical deep Depression Recognition network (OpticalDR). Experiments on CelebA, AVEC 2013, and AVEC 2014 datasets demonstrate that our OpticalDR has achieved state-of-the-art privacy protection performance with an average AUC of 0.51 on popular facial recognition models, and competitive results for DR with MAE/RMSE of 7.53/8.48 on AVEC 2013 and 7.89/8.82 on AVEC 2014, respectively.

Figures (7)

• Figure 1: Different DR approaches: (a) conventional facial recognition with no privacy preservation, (b) facial features-based approaches with limited privacy preservation, and (c) our full privacy-preserving approach that doesn't generate facial images. The DR approaches with regular cameras could be at the risk of sensitive information access by attackers after the images are captured. However, with our approach, no sensitive information would be captured or stored, whether on the client or the server.
• Figure 2: The architecture of the proposed OpticalDR. The optical model comprises a thin lens for image capture in front of the sensor. The deep learning model utilizes the captured image and is jointly optimized with the optical component for depression score recognition.
• Figure 3: ROC curves depicting the performance of facial recognition models under privacy-preserving approaches, including Gaussian blur and defocus methods, alongside our OpticalDR.
• Figure 4: Trade-off between privacy preservation and DR performance among different privacy-preserving strategies.
• Figure 5: The visual effects of (a) the original facial image, (b) Gaussian blur with different sigma values, (c) defocus with varying diameters of bokeh balls, and (d) the optical solution for facial privacy preservation.