Research on the Privacy Security of Face Recognition Technology

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Face recognition technology in applications faces significant privacy risks due to potential data leakage.
• Existing privacy protection methods may not adequately secure sensitive facial data in real-world applications.

Purpose of the Study:

• To propose a novel method based on differential privacy for enhanced privacy security in face recognition applications.
• To protect sensitive parameter information of face recognition models against unauthorized access and recovery attacks.

Main Methods:

• Utilized Bayesian Generative Adversarial Networks (GAN) to generate privacy data with a similar distribution for training.
• Applied differential privacy algorithms to train data, generating privacy-protected labels.
• Developed a lightweight face recognition model ('lightface') to generate noisy tags and employed gradient descent on recovered face feature vectors.

Main Results:

• The proposed method effectively protects face recognition model parameters, significantly reducing the accuracy of recovered images by attackers.
• Analysis of privacy loss provided an accurate privacy protection boundary.
• Demonstrated superior privacy protection capabilities compared to Differential Privacy Stochastic Gradient Descent (DPSGD) and পার্থক্য (PATE).

Conclusions:

• The developed differential privacy-based method offers robust privacy security for face recognition technology in practical applications.
• The approach successfully balances privacy protection with the utility of face recognition models.