Privacy-Preserving Semantic Segmentation Using Vision Transformer.

Area of Science:

• Computer Vision
• Machine Learning
• Cybersecurity

Background:

• Privacy-preserving deep learning methods have primarily focused on image classification.
• Existing methods using encrypted images often suffer from reduced accuracy compared to models trained on plain images.
• Semantic segmentation tasks require robust methods to handle sensitive visual information securely.

Purpose of the Study:

• To propose a novel privacy-preserving semantic segmentation method.
• To leverage the Vision Transformer (ViT) architecture for secure image analysis.
• To maintain high accuracy when processing encrypted images in segmentation tasks.

Main Methods:

• Utilized the Segmentation Transformer (SETR), a ViT-based model, for semantic segmentation.
• Employed image encryption techniques to protect sensitive visual information.
• Introduced a novel embedding strategy within the ViT structure for privacy preservation.

Main Results:

• The proposed method achieves semantic segmentation with encrypted images.
• Demonstrated that the privacy-preserving approach maintains accuracy comparable to using plain images.
• Successfully applied ViT embeddings for privacy-preserving semantic segmentation for the first time.

Conclusions:

• The developed privacy-preserving semantic segmentation method effectively secures sensitive visual data.
• The integration of SETR and image encryption offers a viable solution for accurate segmentation without compromising privacy.
• This work advances privacy-preserving techniques in deep learning for complex computer vision tasks.