ReLU Network Model Extraction

关于

Provides specialized guidance for reverse-engineering the internal parameters of two-layer ReLU neural networks through black-box query access. By leveraging the piecewise linear structure of ReLU functions, this skill facilitates the recovery of weight matrices up to permutation and scaling through techniques like critical point detection, Hessian analysis, and Independent Component Analysis (ICA). It ensures high-fidelity model extraction by emphasizing unique row matching and one-to-one correspondence over simple functional fitting, making it an essential resource for security researchers and developers performing model extraction attacks or neural network reverse-engineering.

主要功能

• Hessian-based and ICA-driven weight recovery methods
• 16 GitHub stars
• Critical point detection for identifying hyperplane boundaries
• Probing techniques to estimate hidden layer dimensions
• Row-wise verification logic to ensure unique parameter matching
• Guidance on avoiding misleading success metrics and numerical errors

使用场景

• Auditing ReLU network transparency and weight configurations
• Reverse-engineering model parameters for security research
• Executing model extraction attacks on proprietary neural network APIs