STFG: Secure and Trustworthy Federated Generative AI

Nitin Bisht is a PhD candidate in Artificial Intelligence at the University of Technology Sydney (UTS), Australia. His research focuses on the statistical foundations of trustworthy AI, with applications in recommender systems and large language models. He develops methods for uncertainty quantification and risk control to address reliability and fairness in personalised recommendation, continual learning and machine unlearning.

His research has been published or accepted at leading international conferences, including ICML, NeurIPS, SIGIR and CIKM, and in IEEE Transactions on Knowledge and Data Engineering. He has also served as a reviewer for major AI and data science conferences, including NeurIPS, ICLR, ICML, AAAI and KDD.

Recommender systems shape what people discover, but the feedback they receive offers only a partial view of what people want. A click may reflect a temporary need, while an item may receive no interaction simply because it was never shown. As systems learn from these signals, uncertain assumptions can influence later recommendations and become reinforced by the feedback they generate. Keeping personalization reliable therefore requires understanding how uncertainty develops through repeated recommendation and learning.

In this talk, I will present statistical approaches to managing this uncertainty. I will begin with adaptive calibration, showing how recommendations can be adjusted as interests change while accounting for limited data and delays in detecting change. Adapting to feedback, however, is only part of the challenge. Recommendations also shape the feedback a system receives, so controlling repeated exposure to similar content can help broaden discovery while preserving relevance. As systems continue learning from this feedback, they must also respect users’ requests to forget information. I will discuss how controls on subsequent updates can limit the risk of that information influencing recommendations again.

Finally, I will discuss the current challenges of extending these statistical approaches to LLM-based recommenders, where interpreting user feedback and repeatedly updating profiles and models introduce further uncertainty.

Chunxu Zhang received her Ph.D. in Computer Science and Technology from Jilin University in 2024 under the supervision of Prof. Bo Yang, and her doctoral dissertation received the university’s Outstanding Doctoral Dissertation honor. Her research interests include federated learning, recommender systems, federated foundation models and graph neural networks.

Her work has been published at leading venues including KDD, WWW, NeurIPS, AAAI, IJCAI and SIGIR, and in IEEE Transactions on Knowledge and Data Engineering. She co-presented a tutorial on personalized federated recommender systems at WWW 2025, where she also served as a session chair.

Recommender systems have become an essential part of modern digital services, enabling personalized access to information, content, and products. Their success has largely relied on centralized learning, where massive user interactions are collected to train shared recommendation models. However, as user data, computational resources, and service environments become increasingly distributed, this paradigm faces new challenges in capturing individual preferences and supporting collaboration across different participants. Federated recommendation offers an alternative that goes beyond keeping user data local. It creates opportunities to rethink how personalization is learned, how knowledge is shared, and how users participate in the recommendation process.

In this talk, I will discuss our research on personalized federated recommendation from this broader perspective. Rather than treating federated learning simply as a distributed implementation of existing recommendation models, we explore how its decentralized nature can enable richer personalization and more flexible collaboration. I will highlight key advances in extending user-specific modeling capabilities, leveraging relationships among distributed users, and incorporating complementary knowledge to improve recommendation. These studies collectively illustrate how federated recommendation can move beyond conventional model sharing toward more effective coordination between individual preferences and collective knowledge.

Finally, I will discuss emerging opportunities and open challenges as recommendation systems evolve toward foundation models, generative AI, and user-centric intelligent services. Particular attention will be given to how personalized models can continuously learn and collaborate in increasingly complex and dynamic environments while respecting user privacy.