Privacy-Preserving CNN Inference for Image Super-Resolution Cross Multiple Ciphertexts

Area of Science:

• Computer Vision
• Cryptography
• Machine Learning

Background:

• Online image super-resolution (SR) services are popular but pose privacy risks due to plaintext image exposure.
• Existing secure CNN inference methods are unsuitable for SR due to the increased output resolution.
• Protecting sensitive image data during SR processing is a critical challenge.

Purpose of the Study:

• To develop a privacy-preserving scheme for Convolutional Neural Network (CNN)-based image super-resolution.
• To enable secure inference for image SR applications without compromising data privacy.
• To improve the efficiency and practicality of secure image SR.

Main Methods:

• A novel multiple ciphertext encapsulation method for secure CNN inference in image SR.
• Design of fundamental homomorphic operations (addition, multiplication, rotation) across ciphertexts.
• Development of a fast, secure upsampling algorithm using pre-weight block masking and cross-ciphertext rotation.
• Implementation of an efficient batched homomorphic 2D convolution with kernel optimization.
• Creation of a polynomial activation function tailored for SR tasks.

Main Results:

• The proposed HE-friendly SR network demonstrates superior performance compared to existing secure solutions.
• The multiple ciphertext encapsulation technique achieves at least a 2x improvement in computational efficiency.
• Significant reduction in memory usage was observed with the new encapsulation method.
• The secure upsampling algorithm offers substantial speedup over direct homomorphic upsampling.

Conclusions:

• The developed scheme effectively addresses privacy concerns in online image SR services.
• The novel techniques provide a practical and efficient solution for secure image super-resolution.
• This work paves the way for more secure and private AI-driven image enhancement applications.