Invertible Privacy-Preserving Adversarial Reconstruction for Image Compressed Sensing

Area of Science:

• Computer Vision
• Image Processing
• Cybersecurity

Background:

• Compressed sensing (CS) algorithms aim to improve image quality.
• High-quality images can inadvertently expose sensitive information in machine recognition.
• Existing CS methods often overlook privacy concerns.

Purpose of the Study:

• To develop an invertible, privacy-preserving adversarial reconstruction method for image CS.
• To balance image quality with data security in CS.
• To enable differential access to reconstructed image data.

Main Methods:

• Proposed a novel adversarial reconstruction technique integrated with CS.
• Images are generated as adversarial samples during reconstruction.
• An additional restoration network is used to recover clean images for authorized users.

Main Results:

• Reconstructed images maintain good visual quality for both semi-authorized and authorized users.
• Semi-authorized users receive adversarial images with significantly reduced recognizability.
• Authorized users can recover sanitized images with recognition performance close to traditional CS.

Conclusions:

• The proposed method effectively enhances privacy in CS image reconstruction.
• It offers a practical solution for scenarios requiring controlled data access.
• The technique balances visual fidelity with robust data protection.