k-Same-Net: k-Anonymity with Generative Deep Neural Networks for Face Deidentification

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Sharing image and video data requires robust privacy protection for personal information.
• Deidentification techniques conceal individual identities in imagery while retaining data utility.
• Existing methods may not offer formal privacy guarantees or sufficient utility preservation.

Purpose of the Study:

• To propose k-Same-Net, a novel face deidentification approach combining Generative Neural Networks (GNNs) and k-Anonymity.
• To provide formal privacy guarantees for deidentified face data within a closed set of identities.
• To enable control over synthetic face generation for tailored deidentification.

Main Methods:

• Developed a GNN capable of generating synthetic surrogate face images.
• Integrated the GNN with the k-Anonymity mechanism for formal privacy guarantees.
• Utilized appearance-related parameters to control synthetic image characteristics (e.g., expression, age, gender).

Main Results:

• Demonstrated the feasibility of k-Same-Net on XM2VTS and CK+ datasets.
• Evaluated deidentification efficacy against state-of-the-art recognition models and competing techniques.
• Showcased utility preservation through facial expression recognition experiments.

Conclusions:

• k-Same-Net offers a viable solution for face deidentification with formal privacy guarantees.
• The approach effectively balances privacy protection and data utility.
• k-Same-Net presents desirable characteristics compared to existing face deidentification methods.