Binxing FANG (Chinese Academy of Engineering)

Title: On AI Agent Safety and AI Safety Fuse

09:15-10:30, Aug 5 (Tue) @Theater 1+2+6

Abstract:

Artificial intelligence agents (such as robots, etc.) have increasingly permeated people's lives. However, if these agents get out of control, how can people prevent it? Under the control of large models, will artificial intelligence agents confront humans? If such a situation really occurs, how should it be dealt with? This article proposes the concept of an "artificial intelligence agent fuse", which is a combination of a "fuse" and a "mantra". The aim is to install the fuse on the AI agent when it leaves the factory. In case of an emergency, the fuse can be "blown" by the "mantra", causing the artificial intelligence agent to lose its power and prevent it from harming humans. This is similar to how an intelligent driving vehicle can prevent harm by applying emergency brakes at critical moments.

Bio:

Binxing Fang, an academician of the Chinese Academy of Engineering, is the honorary dean of the School of Cybersecurity at Guangzhou University, a senior chief scientist at China Electronics Technology Group Corporation, the director of the New Network Research Department at Peng Cheng Laboratory, the chief academic advisor of the School of Computer Science at Harbin Institute of Technology (Shenzhen), and the director of the National Engineering Laboratory for Information Content Security Technology. He is the convener of the Network Security Discipline Evaluation Group of the State Council Academic Degrees Committee and the deputy director of the Network Security Professional Teaching Guidance Committee of the Ministry of Education. Currently, he serves as the president of the Chinese Information Processing Society, the chairperson of the Network and Information Security Technology Committee of the China Standardization Association, the president of the China Cybersecurity Talent Education Forum, and the president of the China Cybersecurity Emerging Technology Security Innovation Forum. He previously served as the president of Beijing University of Posts and Telecommunications, the director of the National Computer Network Emergency Response Technical Coordination Center, and the first president of the China Cybersecurity Association. He has also held positions as the vice president of the China Internet Society, the China Institute of Communications, and the China Computer Federation. He has won six national science and technology progress awards of the first and second class, over ten provincial and ministerial awards, and has authored five books and over 400 articles.

Jianwei LIU (Beihang University)

Title: AI Empowered 6G Network Security: Architecture and Key Technologies

09:15-10:30, Aug 6 (Wed) @Theater 1+2+6

Abstract:

The sixth-generation mobile communication technology (6G) envisions 'land-sea-air-space integrated and intelligent connection of all things,' providing Tbps-level ultra-high speed, sub-millisecond ultra-low latency, and ultra-high reliability communication services for massive application scenarios. However, 6G networks face security challenges including difficulties in ensuring unified management due to heterogeneous interconnection and ubiquitous connectivity, challenges in managing massive devices, and obstacles in establishing network trust.

Leveraging its powerful learning, knowledge reasoning, and perception-decision capabilities, Artificial Intelligence (AI) transcends the limitations of manual expertise. With the AI intelligent engine at its core, deep coordination across all security layers, and adaptive learning through a closed-loop feedback mechanism, an AI-empowered 6G network security architecture is realized. This makes endogenous security within 6G networks possible.

This report focuses on the new requirements and challenges for 6G network security. Its core contents include:

(1) Proposing a deeply integrated security architecture for 6G networks empowered by AI;

(2) Systematically analyzing the core enabling roles and key empowerment points of AI technology in securing the physical layer, network layer, and service layer of 6G;

(3) Comprehensively reviewing cutting-edge research hotspots and key technologies in this field, while providing a forward-looking perspective on future key research directions.

The report aims to offer systematic reference and guidance for deepening research on endogenous security in 6G networks.

Bio:

Jianwei Liu, Professor, Doctoral Supervisor, Founding Dean, National Distinguished Teacher, Expert Receiving State Council Special Government Allowance, School of Cyberspace Science and Technology, Beihang University. He is currently the member of Academic Degrees Committee of the State Council (Subject Assessment Panel); the member of the Teaching Steering Committee for Cyberspace Security, Ministry of Education; the member of the National Teaching Steering Committee for Cryptography Graduate Education; the Standing Council Member of Chinese Association for Cryptologic Research (CACR); the Standing Council Member of China Institute of Command and Control (CICC); the Vice Chair of Cyberspace Security Technical Committee, Chinese Institute of Electronics (CIE); the Vice Chair, Cyberspace Security Technical Committee, China Institute of Command and Control (CICC); the Executive Associate Editor-in-Chief, Journal of Cyberspace Security Science; the Associate Editor-in-Chief, Journal of Network and Information Security; the Senior Member of IEEE. He has authored 16 textbooks, 7 monographs, and 1 translated work, and has led more than 30 national and provincial/ministerial-level research projects, received more than 20 national/provincial teaching awards.

Yang ZHANG (CISPA Helmholtz Center for Information Security) 

Title: Safety Assessment of Large Generative Models

14:00-15:00, Aug 6 (Wed) @Theater 1+2+6

Abstract:

Over the past two years, large generative models have made remarkable advancements, significantly influencing our daily lives. However, recent research highlights serious security and safety concerns associated with these models. In this talk, I will present some of our recent work addressing these challenges. First, I will discuss methods for detecting and attributing machine-generated content. Next, I will examine whether these models are likely to produce unsafe outputs. Finally, I will explore techniques for effectively extracting high-quality prompts from large models’ generated content.

Bio:

Yang Zhang (https://yangzhangalmo.github.io/) is a tenured faculty member at CISPA Helmholtz Center for Information Security, Germany. His research concentrates on trustworthy machine learning including privacy, security and more recenlty LLM safety. Moreover, he works on measuring and understanding misinformation and unsafe content like hateful memes on the Internet. His research has been featured in major media outlets including the Washington Post and New Scientist. He has also received the NDSS 2019 distinguished paper award and the CCS 2022 best paper award runner-up.

09:15-10:30, Aug 7 (Thu) @Theater 1+2+6

Mingxing Zhang (Tsinghua University)

Title: LLM Serving on Heterogeneous Hardware

14:00-15:00, Aug 5 (Tue)  @Theater 1+2+6

Abstract:

Traditional LLM inference is often GPU-centric. However, as GPU utilization nears its limit, cost-effective solutions require broader hardware choices. By leveraging the bandwidth and capacity advantages of diverse GPUs and CPU/DRAM, and by exploiting models’ sequential and sparse characteristics, we can build next-generation architectures. This talk introduces two open-source solutions—Mooncake (KVCache-centric, enabling further operator disaggregation) and KTransformers (CPU/GPU co-inference for sparse models)—both of which significantly reduce LLM serving costs in certain scenarios and are widely adopted in industry.

Bio:

Dr. Mingxing Zhang, Assistant Professor at Tsinghua University, focuses on memory systems research. He is the co-founder of the open-source projects Mooncake and KTransformers. His work has been published in over thirty papers at top international conferences and journals, including OSDI, SOSP, ASPLOS, HPCA, and EuroSys. He has received several prestigious awards, including the FAST Best Paper Award, the SIGSOFT Distinguished Paper Award, and authored the first OSDI paper from a Chinese university. He is a recipient of the ChinaSys Rising Star Award, the Outstanding Doctoral Dissertation Award, and the IEEE TCSC Outstanding Ph.D. Dissertation Award. He previously served as Chief Algorithm Expert and Director of the Innovation Research Institute at Sangfor Technologies, where his work contributed to products used by tens of thousands of clients.