Deep Face Leakage: Inverting High-Quality Faces From Gradients Using Residual Optimization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Collaborative learning trains deep learning models without sharing sensitive participant data, like facial images.
• Gradient inversion attacks can reconstruct private data from model gradients, posing privacy risks.
• Existing attacks show suboptimal face reconstruction, losing critical facial details.

Purpose of the Study:

• To propose DFLeak, an effective approach to enhance face leakage from gradients in collaborative learning.
• To improve the quality and detail of reconstructed facial images from gradients.
• To mitigate privacy risks in facial applications within collaborative learning frameworks.

Main Methods:

• Introduced a superior initialization method for stabilizing the gradient inversion process.
• Integrated prior-free face restoration (PFFR) results into gradient inversion via residual optimization.
• Designed a pixel update schedule to preserve fine facial details and mitigate regularization effects.

Main Results:

• DFLeak achieves more realistic and higher-quality facial image reconstructions.
• The proposed method surpasses the performance of state-of-the-art gradient inversion attacks.
• Enhanced recovery of facial details compared to existing techniques.

Conclusions:

• DFLeak effectively boosts face leakage from gradients, enhancing facial reconstruction.
• The approach strengthens privacy defenses against gradient inversion attacks in collaborative learning.
• DFLeak offers a more robust solution for recovering detailed facial images from gradients.