Recently, Professor GuohuiLi and Dr.MinghuiLi of SSE were respectively granted theNational Natural Science Foundation of China.

The general program of Professor GuohuiLi that is granted the NSFCis“Research on Conversational Recommendation and Interpret ability of Multimodal DataFusion”. Information recommendation is an effective response to information overload. In recent years, multimodal data has experienced explosive growth with large data volume, diversity, and in comprehensibility.With effective treatment and utilization, it will recommend more available knowledge and demonstrate better performance.The utilization of knowledge can alleviate the negative impact of sparsed atain return.In the program,semantic alignment and feature fusion of multi-modal dialogue dataset are studiedto improve the utilization of multi-modal dataset; fine-grained user intention recognition and balancing methods are studied and fused with multi-modalintelligence representation to explore effective recommendation strategies by using similarities between entities;explainablepath reasoning overrecommendation results is studied to improve system transparency, credibility and user satisfaction. The program will help achieve high-quality and explainable recommendations, and is of great significance to promote the widespread application of multi-modal dialogue recommendation systems.

Dr. Minghui Li’s project that is granted the Youth Science Fundof NSFCis "Research on Deep Learning Security Technologies for Image Retrieval Scenarios." Image retrieval based on deep learning is an import antfield for the future development of information retrieval, and has recently attracted widespread attention from the academia.The existing deep image retrieval modelsare confronted with serious security issues, such as private databreaches and vulnerability to adversarial attacks. The project will carry out research on deep learning security technologies for image retrieval scenarios, and explore how to simultaneously ensure privacy, robustness and accuracy of deep image retrieval models.In terms ofdata security, lightweight privacy protection technology in the encryption domain will be studied to improve accuracy and efficiency of privacy-protecting image retrieval solutions.In terms of algorithm security, adversarial sample attack and defense technologies in the image retrieval feature space are studied to improve robustness of the image retrieval model while reducing attack and defense overhead. Finally, deep learning model construction that resists Byzantine attacks in distributed scenariosis studied to improve accuracy and robustness of the image retrieval models as a whole. The project takes the image retrieval scenarios as the starting point to study deep learning security technologies.It is expected to obtain a number of high-quality academic results to provide theoretical and technical support for the development of secure image retrieval and deep learning security.