Local Privacy Protection for Sensitive Areas in Multiface Images

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Face recognition systems face privacy challenges in accurately identifying individuals.
• Existing privacy protection methods may not effectively handle multiface images.
• Goal-driven reasoning offers a novel approach to local privacy protection.

Purpose of the Study:

• To design a goal-driven algorithm for local privacy protection in sensitive areas of multiface images.
• To integrate face recognition, regional growth, and differential privacy within an interactive framework.
• To develop a privacy protection for sensitive areas (PPSA) algorithm that addresses multiface scenarios.

Main Methods:

• Utilized multitask cascaded convolutional network (MTCNN) for face and landmark recognition.
• Employed regional growth seeded by overlapping subgraphs and guided by fusion similarity measurement mechanism (FSMM).
• Implemented a novel privacy budget allocation strategy dividing the total budget (ε) into ε_1 and ε_2 for targeted protection.

Main Results:

• The PPSA algorithm ensures ε-differential privacy for multiface images.
• Noise error in PPSA is independent of image size, unlike the Laplacian algorithm.
• Achieved significant improvements: accuracy (≥16.1%), recall (≥2.3%), and F1-score (≥15.2%) in image classification tasks.

Conclusions:

• The PPSA algorithm offers effective and efficient local privacy protection for sensitive face areas in multiface images.
• The proposed privacy budget allocation enhances the performance of privacy protection in complex scenarios.
• PPSA demonstrates superior performance compared to existing methods, meeting differential privacy requirements.