Enhancing Privacy Protection in Online Federated Learning: A Method for Secure Face Image De-Identification Using a Modified Diffie-Hellman Algorithm

Abstract

The proliferation of face images, alongside their widespread dissemination and easy accessibility through social media, underscores a pressing challenge to personal identification information protection. Conversely, advancements in identity-agnostic computer vision technologies offer valuable benefits, necessitating cautious utilization of face images to safeguard individual privacy. 'Face de-identification', or 'face anonymization', refers to the process of altering an original face image to a nearidentical one that obscures the subject's actual identity. Existing de-identification strategies, despite considerable efforts, often fall short in photo-realism or fail to strike an optimal balance between privacy and utility. This study proposes an approach for generating de-identified facial images using instances, addressing the potential privacy breaches and identity exposure associated with facial features. The proposed system involves a two-stage training process. Initially, a federated learning framework is suggested, enabling knowledge amalgamation through the mutual exchange of model parameters among clients during federated training, devoid of data sharing. Subsequently, sensitive information is secured using an enhanced version of the Diffie-Hellman algorithm coupled with a genetic algorithm. In the event of data loss or corruption, an optimized genetic algorithm (OGA) is employed to successfully restore the data, thereby offering protection against potential insider threats in federated learning. The decryption process is then executed as if the user had initiated the request. Experimental results demonstrate that the proposed federated learning approach delivers performance equivalent to centralized learning, thereby validating the practicality and effectiveness of the suggested architecture. Specifically, a model of the federated learning-deep convolutional neural network (FL-DCNN) achieved an accuracy of 95.2%, precision and F1-score of 95%, recall of 96%, and a final specificity of 96.80%.