Membership inference attack on differentially private block coordinate descent.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models rely on large datasets, often containing sensitive personal information.
• Privacy concerns arise from potential misuse of this data, driving research into privacy-preserving deep learning.
• Differential privacy (DP) is a key technique for protecting data in deep learning models.

Purpose of the Study:

• To analytically evaluate the effectiveness of DP-BCD, a novel privacy-preserving method, against sophisticated privacy attacks.
• To assess the practical privacy preservation capabilities of DP-BCD in both black-box and white-box settings.
• To compare DP-BCD's performance against the state-of-the-art DP-SGD method.

Main Methods:

• Implemented membership inference attacks (MIAs) in black-box and white-box scenarios.
• Evaluated DP-BCD and DP-SGD on benchmark datasets.
• Utilized performance metrics including AUC, attacker advantage, precision, recall, and F1-score.

Main Results:

• DP-BCD demonstrated strong privacy preservation against advanced adversaries.
• The method maintained acceptable model utility compared to existing techniques.
• Experimental results validated DP-BCD's capability to resist sophisticated privacy attacks.

Conclusions:

• DP-BCD effectively preserves the privacy of deep learning models against membership inference attacks.
• DP-BCD offers a promising alternative to DP-SGD with low privacy cost and high accuracy.
• The study confirms DP-BCD's practical viability for safeguarding sensitive training data.