ReLU Model Extraction

소개

This skill provides specialized guidance for reverse engineering black-box neural networks, specifically targeting two-layer ReLU architectures. It outlines robust methodologies for recovering hidden parameters like weight matrices and biases through techniques such as critical point analysis, gradient-based extraction, and activation pattern enumeration. By emphasizing true black-box treatment, the skill ensures that extracted models are verified based on functional equivalence rather than hardcoded implementation details, making it an essential resource for security researchers and machine learning developers working on model stealing, extraction, or auditing tasks.

주요 기능

• Functional equivalence verification through independent output matching
• 16 GitHub stars
• Guidance on avoiding common pitfalls like implementation peeking and hardcoding
• Gradient-based parameter identification for weight matrix recovery
• Strategies for adaptive hidden dimension discovery
• Critical point analysis to identify neuron transition boundaries

사용 사례

• Reverse engineering model parameters for security auditing or model stealing research
• Extracting weights from a black-box neural network when internal access is restricted
• Building functional surrogate models of proprietary ReLU-based networks