Towards Automatic and Scalable Neural Network Verification
Abstract
Although neural networks have achieved remarkable performance in many important tasks, their black-box nature often leads to safetyconcerns in mission-critical applications, such as autonomous control, sensing and decision-making systems. Neural network verification methods are designed to test whether the behavior of a model is provably consistent with certain user-specified properties suchas robustness. Further, training with the bounds provided by neural network verification can lead to networks with certified properties, such as certified robustness. Despite being an important research area in the past four years and having successfully applied to verify small networks (e.g., ACAS Xu airborne collision avoidance models, MNIST/CIFAR models), the current approaches have limited scalability and are not flexible enough to handle general neural network architectures and realistic specifications. Further, current approaches need to re-derive and re-implement the verification methods when handing new architectures, making them hard to applywithout sufficient knowledge in neural network verification. Building on top of the state-of-the-art neural network verification software, Alpha-Beta-Crown, which was developed by the PI s team and won the 2021 International Neural Network Verification Competition (VNN-Comp), the proposal aims to develop novel automatic and scalable neural network verification algorithms to address the above-mentioned challenges. The verification algorithms will lead to an easy-to-use software to enable neural network verification for a wide range of applications. Furthermore, equipped with the new verifiers, a family of theoretically-founded certified robust training methods will be developed to obtain certifiably robust or safe models.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 03, 2023
- Source ID
- N000142312300
Entities
People
- Cho-jui Hsieh
Organizations
- Office of Naval Research
- United States Navy
- University of California, Los Angeles