Reproduce “Efficient Neural Network Robustness Certification with General Activation Functions” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Huan Zhang, Tsui-Wei Weng, Pin-Yu Chen, Cho-Jui Hsieh, Luca Daniel
CROWN introduces a framework to efficiently certify the robustness of neural networks against adversarial attacks, supporting a broad range of activation functions beyond ReLU, including tanh, sigmoid, and arctan. The method achieves up to 26% tighter certified lower bounds for ReLU networks compared to prior approaches while maintaining practical computational efficiency.
Key insights
- Arbitrary activation functions can be effectively bounded by adaptive linear functions, providing a unified approach for robustness certification across diverse network architectures.
- A recursive backward propagation scheme efficiently computes output bounds, extending the capabilities of previous ReLU-specific certification methods.
- Adaptive selection of linear lower bound slopes for ReLU activations leads to tighter certified bounds by minimizing approximation error.
- The framework is scalable, certifying networks with over 10,000 neurons in approximately 1 minute on a single CPU core, making it practical for real-world models.
- The ability to certify various activation functions opens new avenues for research into their intrinsic robustness properties, potentially leading to the design of new, inherently robust network architectures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.