SaTC: CORE: Medium: Collaborative: Towards Trustworthy Deep Neural Network Based AI: A Systems Approach

点击翻译按钮获取中文摘要

Artificial intelligence (AI) is poised to revolutionize the world in fields ranging from technology to medicine, physics and the social sciences. Yet as AI is deployed in these domains, recent work has shown that systems may be vulnerable to different types of attacks that cause them to misbehave; for instance, attacks that cause an AI system to recognize a stop sign as a speed-limit sign. The project seeks to develop methodologies for testing, verifying and debugging AI systems, with a specific focus on deep neural network (DNN)-based AI systems, to ensure their safety and security. The intellectual merits of the proposed research are encompassed in four new software tools that will be developed: (1) DeepXplore, a tool for automated and systematic testing of DNNs that discovers erroneous behavior that might be either inadvertently or maliciously introduced; (2) BadNets, a framework that automatically generated DNNs with known and stealthy misbehaviours in order to stress-test DeepXplore; (3) SafetyNets; a low-overhead scheme for safe and verifiable execution of DNNs in the cloud; and (4) VisualBackProp; a visual debugging tool for DNNs. The synergistic use of these tools for secure deployment of an AI system for autonomous driving will be demonstrated.The project outcomes will significantly improve the security and safety of AI systems and increase their deployment in safety- and security-critical settings, resulting in broad societal impact. The results of the project will be widely disseminated via publications, talks, open access code, and competitions hosted on sites such as Kaggle and NYU's annual Cyber-Security Awareness Week (CSAW). Furthermore, students from under-represented minority groups in science, technology, engineering and mathematics (STEM) will be actively recruited and mentored to be leaders in this critical area. The code for this project will be made publicly available via github.com. Preliminary code for the tools that will be developed is already hosted on this website, including DeepXplore (https://github.com/peikexin9/deepxplore) and BadNets (https://github.com/Kooscii/BadNets/). These repositories will be linked to from a homepage that describes the entire project. The project homepage will be hosted on wp.nyu.edu/mlsecproject.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.