Training Verification-Friendly Neural Networks via Neuron Behavior Consistency (original) (raw)

Authors

• Zongxin Liu Key Laboratory of System Software (Chinese Academy of Sciences) and State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, Beijing, China University of Chinese Academy of Sciences, Beijing, China
• Zhe Zhao RealAI, Beijing, China
• Fu Song Key Laboratory of System Software (Chinese Academy of Sciences) and State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, Beijing, China Nanjing Institute of Software Technology, Nanjing, China
• Jun Sun Singapore Management University, Singapore
• Pengfei Yang College of Computer and Information Science, Software College, Southwest University, Chongqing, China
• Xiaowei Huang The University of Liverpool, Liverpool, United Kingdom
• Lijun Zhang Key Laboratory of System Software (Chinese Academy of Sciences) and State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, Beijing, China University of Chinese Academy of Sciences, Beijing, China

DOI:

https://doi.org/10.1609/aaai.v39i6.32614

Abstract

Formal verification provides critical security assurances for neural networks, yet its practical application suffers from the long verification time. This work introduces a novel method for training verification-friendly neural networks, which are robust, easy to verify, and relatively accurate. Our method integrates neuron behavior consistency into the training process, making neuron activation states remain consistent across different inputs within a local neighborhood. This reduces the number of unstable neurons and tightens the bounds of neurons thereby enhancing the network's verifiability. We evaluated our method using the MNIST, Fashion-MNIST, and CIFAR-10 datasets with various network architectures. The experimental results demonstrate that networks trained using our method are verification-friendly across different radii and architectures, whereas other tools fail to maintain verifiability as the radius increases. Additionally, we show that our method can be combined with existing approaches to further improve the verifiability of networks.

How to Cite

Liu, Z., Zhao, Z., Song, F., Sun, J., Yang, P., Huang, X., & Zhang, L. (2025). Training Verification-Friendly Neural Networks via Neuron Behavior Consistency. Proceedings of the AAAI Conference on Artificial Intelligence, 39(6), 5757-5765. https://doi.org/10.1609/aaai.v39i6.32614

Issue

Section

AAAI Technical Track on Computer Vision V