Defending AI based FinTech Systems against Model Extraction Attacks

该项目旨在开发新的方法，以保护基于人工智能（AI）的金融科技系统免受强大而阴险的模型提取攻击，攻击者可以从系统中窃取AI模型，造成知识产权（IP）侵犯、业务优势中断和经济损失。这可以通过检查各种攻击模型，创建主动和实用的防御，以及在AI模型上发明不可移动的水印来实现。其成果是在部署前保护基于人工智能的金融科技系统的新工具，以及部署后的IP侵权取证工具。通过提高金融科技系统和其他国家关键人工智能系统的安全性，这些能力是有益的。

This project aims to develop new methods for defending artificial intelligence (AI) based FinTech systems from highly potent and insidious model extraction attacks whereby an adversary can steal the AI model from the system to cause intellectual property (IP) violation, business advantage disruption, and financial loss. This can be achieved by examining various attack models, creating active and utility-preserving defences, and inventing non-removable watermarks on AI models. The outcomes are new tools for securing AI-based FinTech systems before deployment and tools for IP violation forensics post-deployment. Such capabilities are beneficial by improving the security and safety of FinTech systems and other nationally critical AI systems.