Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Learn about a novel approach to protecting neural network privacy through formal verification techniques in this 14-minute conference presentation from OOPSLA 2025. Discover how researchers Anan Kabaha and Dana Drachsler Cohen from Technion address the critical challenge of privacy attacks on neural networks that can extract private training data information. Explore the limitations of traditional differential privacy (DP) methods that significantly reduce network accuracy by adding noise for all possible training sets, and understand how Individual DP (iDP) offers a more targeted solution by restricting privacy guarantees to specific training sets. Examine the key innovation of computing iDP deterministic bounds (iDP-DB) to identify inputs that naturally satisfy privacy requirements without noise addition, enabling label-only access with minimal accuracy loss. Delve into the LUCID framework that leverages multiple formal verification techniques including mixed-integer linear programming, hyper-network abstractions, branch-and-bound optimization, and linear constraint encoding to compute the tightest possible privacy bounds. Analyze the impressive experimental results showing LUCID can achieve perfect individual privacy (0-iDP) with only 1.4% accuracy decrease compared to 12.7% reduction from existing DP training algorithms, demonstrated across fully-connected and convolutional networks on four different datasets. Gain insights into cutting-edge research at the intersection of neural network verification, privacy protection, and constrained optimization that advances the field of secure machine learning.
