Provable Verification Bound Towards Convolution Neural Networks Robustness Analysis
Abstract
Deep neural networks have reached unprecedented performance in various tasks such as face recognition and self-driving cars. However, these networks are vulnerable to malicious modification of the pixels in input images, known as adversarial examples. This raises serious concerns if one would to rely on such a recognition model in critical applications where classification error may be life-threatening. In view of the threat posed by adversarial examples, how to protect neural networks from being tricked by adversarial examples has become an emerging research topic. Previous studies of defense against adversarial examples may protect the network from certain adversarial examples, but there is only empirical evidence that they do so, and the robustness of the network is not guaranteed since it is impossible to train or evaluate all possible adversarial examples. In view of the fact that convolutional neural network (CNN) has gained much importance in computation vision and image classification tasks, while previous research on the robust verification of CNNs is rather limited, in this proposal we focus on the robustness analysis for neural networks with CNN backbone.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 16, 2024
- Source ID
- FA23862314068
Entities
People
- Pei-Yuan Wu
Organizations
- Air Force Office of Scientific Research
- National Taiwan University
- United States Air Force