Authors: Qiushi Li, Yan Zhang, Ju Ren, Qi Li, Yaoxue Zhang

Published on: April 05, 2024

Impact Score: 7.6

Arxiv code: Arxiv:2404.04098

Summary

• What is new: A privacy-preserving framework, VisualMixer, that uses pixel shuffling to protect visual DNN training data without injecting noises.
• Why this is important: Existing privacy protection techniques such as Differential Privacy cannot efficiently protect visual features of image datasets.
• What the research proposes: VisualMixer protects visual privacy by determining regions for pixel shuffling according to the Visual Feature Entropy metric, shuffling pixels without injecting noise.
• Results: VisualMixer effectively preserves visual privacy with an average accuracy loss of only 2.35 percentage points, without significant performance degradation.

Technical Details

Technological frameworks used: VisualMixer

Models used: Deep Neural Networks (DNN)

Data used: Real-world datasets

Potential Impact

Companies in autonomous driving and medical imaging could greatly benefit, whereas existing privacy protection technology providers may face disruption.

Want to implement this idea in a business?

We have generated a startup concept here: SecureVision.