Speech Title: AI-empowered Trust and Trustworthy AI

Abstract: While Artificial Intelligence (AI) is contributing to the advancement of human society, it also presents us with new challenges. Its trustworthiness is worthy of in-depth exploration. This talk elucidates the aids of AI for trust and indicates the problems of AI’s trustworthiness, especially potential attacks on AI and the factors that impact AI trust. I will summarize various attacks suffered by AI in its life cycle. In particular, I will introduce recent research achievements of my team, including LLM-aided trust evaluation, GNN-based robust and context-aware trust evaluation models, Large Language Model’s trust problems, a stealthy and practical audio backdoor attack with limited knowledge, LLM enhanced GNN robustness against poisoning attacks. Finally, several insights are proposed regarding AI trust management, especially an agentic system’s trust management.

Speech Title: Lightweight Semantic Video Communications for Remote Driving

Abstract: The explosive growth of multimedia data, connected devices, and real-time intelligent applications is creating unprecedented challenges for existing communication infrastructures. Traditional systems that transmit raw or compressed data often suffer from high latency and inefficient use of bandwidth, particularly in delay-sensitive applications such as remote driving. Semantic communications offer a promising paradigm shift by focusing on conveying meaningful information rather than transmitting data itself. This talk introduces a novel low-latency video semantic communication framework for remote driving, based on an asymmetric encoder–decoder architecture. The framework extracts and transmits only essential semantic features, while generative AI at the receiver reconstructs high-quality, semantically consistent video. We also present a prototype implementation that integrates semantic feature extraction, efficient transmission, and AI-based video reconstruction, demonstrating the practical feasibility of the proposed approach. The talk further discusses the broader vision of integrating semantic communications, generative AI, and intelligent networking toward task-oriented, meaning-aware, and resource-efficient 6G systems, highlighting key challenges and future research opportunities.

Speech Title: "Don't Ask Carelessly!"--a preliminary study on security and privacy protection for LLM prompts

Abstract: Large language models have become powerful general-purpose interfaces, but careless prompting may expose users to serious privacy and security risks. A prompt may contain sensitive personal or confidential information, while malicious or toxic prompts may also induce unsafe model behavior through jailbreaking. This talk presents a preliminary study on protecting the security and privacy of LLM prompts from three perspectives. First, for text prompts, we study lightweight prompt privacy sanitization that can run locally on the user side. The key idea is to jointly measure the importance and privacy risk of each word, so that sensitive but task-irrelevant information can be replaced while preserving answer quality. Second, we extend the discussion to vision-language models, where uploaded images may contain private regions that are unnecessary for answering the user’s question. We introduce a utility-aware image sanitization strategy that protects private visual information while maintaining the accuracy of multimodal responses. Third, we examine jailbreak prompts through the lens of subtoxic questions and gradual attitude change, proposing the idea of a values shadow model for low-cost local evaluation and optimization of jailbreak-prompt strength. Together, these studies highlight the need for prompt-level protection mechanisms that are accurate, lightweight, interpretable, and compatible with existing LLM services.