Although false information detection models excel in terms of accuracy, they often suffer from issues of "overconfidence" or "lack of confidence" in their prediction confidence levels, significantly compromising their reliability in practical applications. When false information spreads rampantly on social media, it becomes crucial to construct a credible detection system that balances accuracy and reliability. Recently, Sun Rui (first author), an undergraduate student from the Class of 2021 at Xuancheng Campus of our school, under the guidance of Associate Professor Hu Wenbo (corresponding author), proposed an innovative framework called RATE (Robust Adversarial Training and Temperature-scaled Ensemble Framework). The paper based on this framework was successfully selected for the International Conference on Multimedia Retrieval (ICMR) 2025, a premier event in the field of multimedia retrieval. This achievement stems from the theoretical foundation and scientific research literacy that Sun Rui accumulated through his participation in the "Hubing Machine Learning Reading Group" during his sophomore year. Through systematic training in classic literature reading and project practice, Sun Rui gradually developed his research capabilities in multimodal false information detection. His work breaks through the bottleneck of confidence calibration in existing models, providing a new credible AI paradigm for the governance of false information on social media. This fully demonstrates the educational effectiveness of our school's "integration of science and education" training model.

Paper Title：RATE: Robust Adversarial Training and Temperature-scaled Ensemble Framework for Trustworthy Misinformation Detection

Authors ：Rui Sun, Wenbo Hu, Qiang Liu, Richang Hong

Figure 1.  RATE Network Framework Structure

Figure 2.  Example of Confidence Calibration for Fake News Detection

ICMR is a series of conferences under the auspices of the Association for Computing Machinery (ACM), serving as a premier international conference in the field of multimedia retrieval. It focuses on innovative research in areas such as multimedia retrieval, cross-modal learning, and content understanding. The conference was formed in 2011 through the merger of the long-established Conference on Image and Video Retrieval (CIVR) and the International Conference on Multimedia Information Retrieval (MIR). Currently, ICMR is recommended as a B-class conference by the China Computer Federation (CCF).