Privacy-Preserving Image Classification Using ConvMixer with Adaptative Permutation Matrix and Block-Wise Scrambled...

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Conventional privacy-preserving image classification methods often require adaptation networks, increasing computational costs, especially for large images.
• Existing block-wise scrambled encryption techniques face challenges in balancing privacy with computational efficiency.

Purpose of the Study:

• To propose a novel privacy-preserving image classification method that eliminates the need for an adaptation network.
• To enable the use of block-wise scrambled images with ConvMixer for both training and testing, enhancing efficiency and accuracy.
• To ensure strong robustness against various attack methods while maintaining high classification performance.

Main Methods:

• Developed a modified ConvMixer architecture capable of directly processing block-wise scrambled images.
• Integrated block-wise scrambling with the ConvMixer model for privacy-preserving image classification.
• Evaluated computational costs against state-of-the-art privacy-preserving deep neural networks (DNNs).

Main Results:

• Achieved high classification accuracy on CIFAR-10 and ImageNet datasets using block-wise scrambled images.
• Demonstrated strong robustness against ciphertext-only attacks.
• Confirmed significantly lower computational resource requirements compared to existing methods.

Conclusions:

• The proposed method offers an efficient and effective solution for privacy-preserving image classification.
• Eliminating the adaptation network reduces computational overhead, making it suitable for large-scale applications.
• The technique provides a strong balance between privacy, accuracy, and computational efficiency in deep learning models.