A Quaternion Rotation-Enhanced Differential Privacy Framework for Image Privacy Protection

Area of Science:

• Computer Vision
• Machine Learning
• Data Privacy

Background:

• Deep learning models process sensitive user data, posing privacy risks.
• Existing privacy-preserving methods for image data lack balance between privacy and accuracy.
• Limited adaptability of current schemes across diverse visual tasks.

Purpose of the Study:

• To develop a novel privacy-preserving scheme that balances high accuracy and robust privacy protection.
• To introduce a dedicated neural network architecture for processing encrypted visual data.
• To improve the adaptability of privacy-preserving methods for various computer vision applications.

Main Methods:

• Proposed quaternion rotation-enhanced differential privacy (QRDP) to encrypt RGB images into quaternion space.
• Implemented differential privacy (DP) by adding noise to encrypted representations for theoretical privacy guarantees.
• Developed the quaternion Fourier Transformer (QFT) network with specialized modules for encrypted image feature extraction.

Main Results:

• QRDP demonstrated strong security against various attacks, confirmed by image quality assessments.
• The QFT network achieved outstanding performance in encrypted image classification and cross-modal retrieval tasks.
• The proposed QRDP scheme outperformed existing methods in both privacy protection and model accuracy.

Conclusions:

• QRDP offers a secure and effective solution for privacy-preserving deep learning with visual data.
• The QFT network successfully extracts robust features from encrypted images, enabling diverse applications.
• This research advances the field of privacy-preserving machine learning for sensitive image data.