Effects of JPEG Compression on Vision Transformer Image Classification for Encryption-then-Compression Images

Area of Science:

• Computer Vision
• Image Processing
• Cryptography

Background:

• Image classification in the encrypted domain is crucial for privacy preservation.
• Previous methods achieved high accuracy but did not consider compression effects.

Purpose of the Study:

• To evaluate the impact of JPEG compression on encrypted image classification using Vision Transformer (ViT).
• To investigate the trade-off between data reduction and classification accuracy for compressed encrypted images.

Main Methods:

• An encryption-then-compression system was used for test images.
• A Vision Transformer (ViT) model was trained on plain images.
• Encrypted test images were compressed using JPEG lossy compression.
• Classification accuracy was verified on compressed encrypted images.

Main Results:

• JPEG compression significantly reduces the data size of encrypted images (over 90% reduction with quality factor 85).
• Classification accuracy is highly preserved, maintaining over 98% accuracy.
• JPEG compression effectiveness was confirmed through comparison with linear quantization.

Conclusions:

• Compressed encrypted images can be classified with high accuracy, comparable to uncompressed encrypted images.
• This study demonstrates the first successful classification of JPEG-compressed encrypted images without significant accuracy loss.
• The findings enable efficient and private image classification through significant data reduction.