Encrypted Image Classification with Low Memory Footprint Using Fully Homomorphic Encryption.

Area of Science:

• Computer Science
• Cryptography
• Machine Learning

Background:

• Deep Neural Networks (DNNs) excel at image classification but raise privacy concerns for sensitive data.
• Current methods often overlook the privacy implications of training and deploying models on personal or confidential images.
• There is a need for secure computation techniques that protect data during machine learning processes.

Purpose of the Study:

• To develop a privacy-preserving image classification system using machine learning and cryptography.
• To implement a Residual Network (ResNet) model capable of processing encrypted images.
• To ensure that only the intended user can decrypt and view the classification results.

Main Methods:

• Exploration of the intersection between Machine Learning and cryptography, specifically Fully Homomorphic Encryption (FHE).
• Development of a Residual Network architecture optimized for FHE computations.
• Implementation using the Cheon-Kim-Kim-Song (CKKS) scheme for approximate encrypted computations.
• Circuit design to minimize memory requirements and computational overhead.

Main Results:

• A novel FHE-based Residual Network for encrypted image classification was proposed and implemented.
• The proposed circuit significantly reduces memory usage compared to prior works.
• An encrypted ResNet20 model achieved 91.67% accuracy on the CIFAR-10 dataset in under five minutes on a laptop.
• Achieved accuracy is comparable to the plain model's accuracy (92.60%).

Conclusions:

• Fully Homomorphic Encryption (FHE) can be effectively combined with Deep Neural Networks for privacy-preserving image classification.
• The developed system offers a practical solution for analyzing sensitive image data without compromising privacy.
• The method demonstrates a viable trade-off between computational efficiency, memory usage, and classification accuracy.