City University of Macau has become the only two private universities supported by the National Natural Science Foundation of China in Hong Kong and Macau since 2024. Dr. Lefeng Zhang and Dr. Chi Liu , young teachers from the Faculty of Data Science, have been awarded the National Natural Science Foundation of China Youth Project. The approved projects are "Research on Forgetting and Privacy Balance of Non parametric Verifiable Machines for Large Models" and "Research on Defense Techniques for AI Image Synthesis in Complex Open Interactive Environments".
Project Introduction (Lefeng Zhang)
The improvement of computing power and model architecture has led to the development of modern artificial intelligence models towards large-scale and complex models. However, the application and popularization of large-scale models have brought a series of challenges to information security, content review, and legal regulation. Machine forgetting aims to remove specific information contained in the model and is an important means of managing and constraining large models.
The complex structure and massive parameters of large models make traditional forgetting strategies that rely on parameter modification no longer efficient or even ineffective; The model changes brought about by information removal pose potential privacy leakage risks to users; The lack of interpretability within the model makes machine forgotten authentication more difficult.
We plan to use knowledge filtering and relative entropy techniques to construct a parameter free machine forgetting algorithm for machine forgetting in large models, in order to eliminate the dependence of machine forgetting on model parameters; Propose privacy protection strategies for machine forgetting using game theory and Nash equilibrium, achieving a balance between forgetting and privacy protection; Using the influence function and gradient matching mechanism to find conjugate data, design a trustworthy machine forgetting authentication scheme for users.
The relevant research results have enriched the theory and method system of machine forgetting, which can be used in fields such as information filtering and model updating that require rapid removal of information from the model, providing methods and technical support for protecting user rights and achieving large-scale model supervision.
Project Introduction (Chi Liu)
The rapid development of AI image synthesis technology has greatly enriched the creation of digital content, but its emerging interaction modes with complex open interaction environments have brought unprecedented information security and digital copyright risks.
At present, defense technologies around AI image synthesis cannot meet the differentiated defense needs in open interactive environments. Therefore, this project will focus on the research of new "defense analysis, defense implementation, defense verification" system level defense technologies for AI image synthesis in open interactive environments. It will construct a user centered, security and availability balanced, graded mechanism, and proactive defense strategy to enhance the security protection of AI image synthesis applications in real application scenarios, solve the key challenges of AI image synthesis defense technology in real open interactive environments, and provide new strategies for information security management and digital copyright in the AI era.
Participate in the Hong Kong and Macau Youth Science and Technology Talent National Conditions Study Course
Assistant Professors Zhang Lefeng and Liu Chi will represent the City University of Macau as awarded with the National Science and Technology Youth Project and participate in the Hong Kong Macau Youth Science and Technology Talent National Conditions Study Course jointly organized by the China Association for Science and Technology and the National Natural Science Foundation of China.
