DPO-Face: Differential privacy obfuscation for facial sensitive regions

Abstract

User-sensitive face images captured by widely used image-collection devices are frequently shared on social media. If these images are misused, they may pose a serious threat to users’ privacy. To ensure both privacy-preserving and image usability, this work introduces a Differential Privacy Obfuscation method of Face images (DPO-Face), which addresses the current limitations in balancing privacy and recognition accuracy. DPO-Face effectively balances privacy preservation and recognition accuracy to meet practical application demands. First, sensitive and non-sensitive regions of the image are accurately identified and located using an improved hybrid convolutional neural network by DPO-Face. Subsequently, face parsing technology is employed to precisely segment the input face image into multiple internal and external facial components. Moreover, precisely adjusted noise is introduced to the internal facial component regions using a differential privacy mechanism to disturb them, effectively protecting the privacy information of these regions while leaving the non-sensitive external components unchanged. Finally, the privacy-protected image is transmitted to the face detection and recognition module to evaluate the effectiveness of the privacy protection, such as maintaining high face detection and recognition accuracy. Experimental results demonstrate that DPO-Face meets $\varepsilon$-local differential privacy requirements, achieving recognition rates of 91%–96% and a maximum privacy protection success rate of 0.9720. This method allows the privacy level to be precisely adjusted, preventing privacy leaks to honest but curious third-party servers, thus achieving a balance between privacy-preserving and usability.