Towards Trustworthy Large Language Model Systems
Large Language Models (LLMs) have become a cornerstone of modern artificial intelligence, offering unprecedented capabilities in natural language processing and generation. Their dependability, however, is a multifaceted concept that encompasses reliability, security, and privacy. Various recent studies have illustrated rather low dependability in LLM systems, illustrating concerns that jeopardize real-world usage of LLMs. In this talk, Shuai will introduce their several recent works on assessing and enhancing the dependability of LLMs, agents, and their underlying systems. He will also discuss some promising future directions that facilitate building reliable LLM-integrated systems and data privacy systems.
Shuai Wang is an Associate Professor at the Department of CSE, HKUST. He is broadly interested in computer security, with a particular focus on AI security, data privacy, and software security. He received the Early Career Award from the Hong Kong RGC in 2020, several industrial research awards from Google/Alibaba/Tencent, as well as several best paper awards and ACM SIGSOFT Distinguished paper award.
From Gigabytes to Megabytes: Overcoming the Excessive Space Usage of Modern Zero-knowledge Systems
Zero-knowledge succinct non-interactive arguments (zkSNARKs) are widely considered for adoption in applications such as scaling and bridging blockchains, or verifiable machine learning. Unfortunately, modern zkSNARKs are notorious for their excessive prover space requirements. That is, producing a proof may require orders of magnitude more space than natively performing the computation itself, which makes their use for large-scale problems prohibitive. In this talk, we propose the first practical space-efficient zkSNARKs that address this issue by introducing novel building blocks that require significantly less prover space, without critical sacrifices to its runtime. In particular, we propose the first sumcheck protocol with optimal prover time in the streaming setting, and a novel polynomial commitment scheme that outperforms all prior works in prover time (and has tunable space requirements). We implemented and benchmarked our schemes across multiple applications, such as (a) inference of pruned Multi-Layer Perceptron, (b) batch AES128 evaluation, (c) select-and-aggregate SQL queries, and (d) verifiable decision tree training. Our experiments demonstrate space-reduction of up to 240x compared to prior state-of-the art works, e.g., when training a decision tree over a dataset of size 400MB, our overall prover space usage is 560MB—less than 1.4x that of just storing the dataset. Surprisingly, this space reduction is also combined with faster prover runtimes by 3x-56x, making our schemes ideal candidates for practical adoption.
Dimitrios Papadopoulos is an associate professor with the Computer Science and Engineering Department of the Hong Kong University of Science and Technology. Prior to this, he received his Ph.D. in computer science from Boston University and was a Post-doctoral researcher at the Institute for Advanced Computer Studies at the University of Maryland. His research is focused on the development of cryptographic protocols for verifiable computation, zero-knowledge proofs, secure computation, encrypted databases, oblivious computation, and other applications. He has published numerous papers in top-tier venues spanning computer security, theoretical and applied cryptography, and database security.
Using AI, a double-edge sword, in Cybersecurity World
Artificial Intelligence (AI) has become a pivotal force in cybersecurity, acting as both a shield and a weapon. Organizations leverage AI for threat detection, log analysis, and behavioral anomaly detection to defend against attacks. Conversely, adversaries exploit AI to automate phishing, craft deepfake scams, and evade traditional security measures.
This talk examines AI’s dual role, highlighting its success in automating SIEM systems, enhancing endpoint protection, and predicting zero-day vulnerabilities. We will also address emerging risks, such as AI-generated malware, adversarial attacks that fool machine learning models, and privacy concerns stemming from pervasive monitoring. By analyzing real-world cases—from AI-augmented penetration testing to AI-powered ransomware—we underscore the need for robust, ethical frameworks to govern AI’s use. The session concludes with strategies to harden AI systems against misuse while harnessing their potential for proactive cyber defense.
Ricci Ieong is the founder and Principal Consultant of eWalker Consulting (HK) Ltd. as well as Adjunct Assistant Professor in the Department of Information Systems, Business Statistics and Operations Management of HKUST, Computer Science and Engineering Department of HKUST and part-time lecturer for the Master of Financial Technology in CUHK.
Ricci has over 25 years of industry experience in the Information Technology Industry as well as more than 20 years of experience in IT Security area specialized in Security Risk Assessment, IT Audit, Ethical Hacking & Penetration Test, Smart Card & Biometrics System deployment and Computer Forensics Investigation.
Ricci first started his IT security career in the CyberSpace Center of HKUST in 1997 after he obtained his MPhil degree from COMP (HKUST). Then in 2000, he founded the first Penetration Test Center within Hewlett-Packard. After working in HP for 5 years, he founded his own IT security company where he leads the IT security planning, IT security assessment, Digital Forensics Investigation, Penetration test and IT audit project as well as security management solution design projects.
Apart from running his consulting business, Ricci delivers lectures in local universities. He is both an Adjunct Assistant Professor teaching Cybersecurity courses and an authorized trainer in AWS Academy in Hong Kong University of Science and Technology (HKUST).
Ricci is the founding member and the Vice Chairman of professional development of Cloud Security Alliance (HK & Macau Chapter) and has served on CSA Cloud Incident Response Working Group and Certificate of Cloud Auditing Knowledge (CCAK) Working Group. He is also the founding member and council member of Information Security and Forensics Society of Hong Kong. He is an active speaker at numerous security events, including CSA summits, in Hong Kong and throughout APAC. Since then, he participated in conducting cloud security training including securing private and public cloud environment as well as Certificate of Cloud Security Knowledge (CCSK) training. He also activate participate in promoting cloud computing and driving cloud security in Hong Kong market. He is one of the recipients of 2021 Ron Knode Service Award awarded by CSA.
Detecting Machine-Generated Fake News Using LLM-Driven Rationales and Adversarial Contrastive Learning
The advent of large language models (LLMs) has revolutionized online content creation but also facilitated the proliferation of low-cost, machine-generated fake news. Effective detection of such content is crucial and requires a robust machine learning model that can detect both the authenticity (legitimate versus fake) and the authorship (human versus machine) of online content. Without this capability, there is a substantial risk of misclassifying machine-generated fake news as legitimate human content and vice versa. Furthermore, knowing the authorship of fake news helps online platforms understand its impact on user engagement and informs targeted interventions and strategic decisions. To address the challenges associated with detecting fake news amplified by LLMs, we propose an innovative detection framework that harnesses the internal reasoning capabilities of LLMs and employs an enhanced adversarial contrastive learning model. This framework is meticulously designed to differentiate between human-generated legitimate, human-generated fake, machine-generated legitimate, and machine-generated fake news. We evaluate our framework using MegaFake, a comprehensive, theory-driven, open-source dataset that includes both human and machine-generated legitimate and fake news articles. Our systematic evaluation demonstrates that the proposed framework outperforms existing detection methods. Our study offers significant research and practical insight for the detection of fake news in the era of LLM.
Dr. Boris Ka Chung Ng is an Assistant Professor and Presidential Young Scholar in the Department of Management and Marketing, Faculty of Business, at The Hong Kong Polytechnic University. He holds a PhD in Information Systems and a BSc in Risk Management and Business Intelligence from The Hong Kong University of Science and Technology. Dr. Boris Ng’s research stands at the intersection of multidisciplinary fields and leverages advanced machine learning techniques to address critical societal challenges and generate business insights. His work contributes to the sustainable development of AI in business and society, focusing on two primary areas: (1) FinTech & AI and (2) Computational Social Science. His research offers both theoretical and practical implications for academia, institutions, and policymakers. His work appears in leading business journals, including Journal of Management Information Systems and Production and Operations Management, and has been presented at premier Information Systems conferences.
Managing Algorithm Manipulation through Human-AI Collaboration
Machine learning models are widely used in critical business decisions such as hiring, marketing and financing, but there is a growing concern that strategic individuals may manipulate their features to obtain favorable outcomes from the machine learning models. A widely discussed solution is to involve humans in overseeing the machine decisions, but it remains unclear when and how to engage humans in different decision situations. This paper develops an economic model to analyze the impact of feature manipulation on decision dynamics and how human-in-the-loop affects such dynamics. We have two key interesting findings: First, increased feature manipulation does not always make decision-makers more selective. For example, while moderate resume manipulation may lead a firm to have a higher standard for accepting job applicants, a high level of manipulation may lead the firm to lower the accepting standard. Second, while feature manipulation always reduces a firm's payoff without human-in-the-loop, it may increase a decision maker's payoff with human-in-the-loop, suggesting human involvement may turn feature manipulation into desirable activities. Our findings have important implications for the practice of involving human-in-the-loop to combat algorithm manipulation.
Jiali Zhou is an assistant professor at American University in Washington, D.C. His main research interest is cybersecurity economics and policies, such as security crowdsourcing, hacker forums governance and AI risks. He is a recipient of ACM SIGMIS Doctoral Dissertation Award and his research won awards in conferences such as the best student paper award in WISE.
Merchants of Vulnerabilities: How Bug Bounty Programs Benefit Software Vendors
This paper uncovers an important paradox in software security: bug bounty programs (BBPs), ostensibly created to enhance software security, perversely lead vendors to release less secure software. Using a game-theoretic model to investigate the economic implications of BBPs, we show that vendors strategically reduce pre-release testing and rely instead on BBPs for post-release management of software vulnerabilities (SVs). This behavior emerges because BBPs provide vendors greater assurance that severe SVs can be privately reported and patched, thereby reducing the perceived risk of uncoordinated public disclosures. In addition to this primary finding, our analysis examines the strategic interactions and trade-offs among software vendors, ethical hackers (white-hat hackers), and malicious hackers (black-hat hackers), revealing several non-obvious insights. First, we demonstrate that participation in BBPs can enhance software vendors' expected profits when the benefits to ethical and malicious hackers, adjusted for effort costs, are comparable. This explains the selective adoption of BBPs and their varying success. Second, we show that offering higher bounties incentivizes ethical hackers to exert greater effort, increasing the likelihood that they discover severe vulnerabilities before malicious hackers. However, this introduces trade-offs related to costs, software release timing, and the reputational benefits derived by ethical hackers, thereby complicating the vendor’s decision on bounty sizes. Third, we find that the optimal number of ethical hackers to invite into a BBP depends on the expected number of malicious hackers targeting the software; interestingly, this optimal number is always lower than, but increases with, the malicious hacker count, providing practical guidelines for program design. These findings challenge the conventional view of BBPs as purely post-release security tools, illustrating their role in reshaping vendor incentives toward accelerated—albeit riskier—software launches. The study underscores the importance of balancing speed-to-market with robust patching practices and transparent disclosure, and urges policymakers to consider regulatory frameworks that address these emergent trade-offs in cybersecurity strategy.
Muhammad “Zia” Ziauddin Hydari is an Assistant Professor of Business Administration at the University of Pittsburgh. He holds a PhD in Industrial Administration (Business Technologies) from Carnegie Mellon University, an MS in Engineering and Management from MIT, an MS in Computer Science from UIUC, and a BEng in Computer Engineering from NED University (Karachi). Before academia, Zia was Principal Product Manager at Oracle USA and a management consultant in telecom and pharmaceuticals. His research streams are in healthcare technologies and cybersecurity. His work has appeared in Management Science, Annals of Emergency Medicine, and Communications of the ACM. He’s presented at CHITA, ICIS, WISE, INFORMS, the POMS Conference, and NBER’s Economics of IT Workshop.
Cybersecurity as a Credence Good: Market Impact and Policy Solutions
The cybersecurity services market is characterized by significant information asymmetry between experts and clients, making it a classic case of a credence good. This asymmetry enables experts to exploit client uncertainty by inflating perceived risks to increase profits. We develop an economic model to illustrate how such dynamics lead to welfare losses, even when clients can perform noisy self-assessments of risk. To address this, we propose two incentive-compatible mechanisms—an additive and a multiplicative scheme—to promote truthful risk disclosure by experts. We analyze the conditions under which these mechanisms are effective, and discuss their applicability across various cybersecurity contexts, while noting potential trade-offs in budget balance. Our findings offer actionable insights for policymakers and regulators aiming to improve market efficiency, strengthen cybersecurity outcomes, and enhance social welfare.
Ping Fan Ke is an Assistant Professor of Information Systems in the School of Computing and Information Systems, at Singapore Management University. His research interests include economics of cybersecurity and blockchain, and application of generative AI. He has published research in various information systems journals, including Information Systems Research, Production and Operations Management, and ACM Transactions on Management Information Systems. He received his Ph.D. in Information Systems from the Hong Kong University of Science and Technology in 2018.
Negative Reputation Spillover in Sensitive Product Markets: Evidence from a Hacker Marketplace
The high anonymity and weak regulation render sensitive product marketplaces attractive arenas for low-quality sellers. Negative feedback can help differentiate these sellers, but it may also generate spillover effects on other sellers. This paper investigates the spillover effect of negative reputation in sensitive product markets and the effect of sellers’ social interaction on such spillover effects. By analyzing negative feedback and transactions in a market for hacking-related products, we find that negative feedback not only decreases sales for the targeted seller but also creates a negative spillover effect, reducing sales for sellers offering similar products without negative feedback. This finding of negative spillover sharply contrasts with the positive spillover typically found in ordinary product markets. We also find that while most sellers on the periphery of a forum social network are susceptible to the spillover effect, a few central sellers remain unaffected, implying a two-tier economy within the marketplace. Further analysis suggests that social networking activities enable hacker sellers to establish relationships with potential buyers (relationship channel) and signal their trustworthiness (information channel), thereby mitigating the negative spillover effect. We discuss the implications for intervention in sensitive product markets.
Jiali Zhou is an assistant professor at American University in Washington, D.C. His main research interest is cybersecurity economics and policies, such as security crowdsourcing, hacker forums governance and AI risks. He is a recipient of ACM SIGMIS Doctoral Dissertation Award and his research won awards in conferences such as the best student paper award in WISE.
Seung Hyun Kim is a Professor of Information Systems (with YSB Research Chair Professorship) at the School of Business, Yonsei University. He received his Ph.D. and M.S. from the Carnegie Mellon University, and his bachelor’s degrees from the Yonsei University. His primary research interests include economics of information security and privacy, mobile commerce, online platform, healthcare IT, and digital marketing. His work has been published in leading academic journals including MIS Quarterly, Information Systems Research, IEEE Transactions on Engineering Management, Decision Support Systems, Information & Management, Journal of Interactive Marketing, and Communications of the ACM. He is currently serving as an Associate Editor for MIS Quarterly.
Kai-lung Hui is the Acting Dean of HKUST Business School. He is currently the Elman Family Professor of Business and the Chair Professor of Department of Information Systems, Business Statistics and Operations Management. He was an Associate Dean for Undergraduate Studies (2013-2015) and Research (2019-2021) of the HKUST Business School. Before joining HKUST, he was with the City University of Hong Kong (2006-2008) and National University of Singapore (2000-2008). He has taught undergraduate, MSc, MBA, EMBA, PhD, DBA and executive courses in information privacy and security management, Fintech and blockchain, business analytics, electronic commerce, business strategy, and technology management, among others He has provided expert advice and consulting services to various government and non-government organizations including the Insurance Authority of Hong Kong, Intellectual Property Department of the Hong Kong SAR Government, World Intellectual Property Organization (WIPO), and Ministry of Law of Singapore. He is currently an honorary advisor of the Hong Kong Police College and is serving as an advisor for UC.NOW and Insurance Authority of Hong Kong. He frequently speaks in conferences and forums on digital piracy, cyber security, privacy, fintech, and technology policies, including those organized by WIPO/OECD, Office of the Privacy Commissioner for Personal Data, ISACA, Insurance Authority, Communications Association of Hong Kong, and HKUST. He currently serves as an Editor for Journal of Management Information Systems, and was the Senior Editor for Information Systems Research, which are two of the top three information systems journals used in the Financial Times’ business school ranking.
Breached and Denied: The Cost of Data Breaches on Individuals as Mortgage Application Denials
While a large body of the information systems (IS) literature has investigated the antecedents and consequences of data breaches in organizations, we do not have a good understanding of whether a data breach has a material impact on individuals whose private information is compromised and how much damage it causes. We overcome empirical challenges in investigating the impact of data breaches on individual victims by utilizing a unique natural experimental setting that allows us to credibly identify treated and controlled populations—the breach of South Carolina (SC) taxpayer records in 2012. With residents in SC as the treatment group and those in Georgia and North Carolina as the control group, our difference-in-differences estimations find that after the breach to the SC Department of Revenue, there was a significant increase in denials to SC residents’ residential mortgage applications for refinance and home improvement. We also find that the adverse impact of the breach was more profound for Black and Hispanic residents. Our study provides significant theoretical and policy implications with respect to the harm and costs of a large-scale data breach.
Min-Seok Pang is Karen A. and William S. Monfre Professor in Business and a Professor of Information Systems and Analytics at Wisconsin School of Business, University of Wisconsin-Madison. He has received a B.S. in Industrial Engineering and an M.S. in Management from Korea Advanced Institute of Science and Technology (KAIST) and holds a Ph.D. in Business Administration from University of Michigan. He was previously a faculty member at Temple University and George Mason University. His research interests include, among others, cybersecurity management, strategic management of information technologies, and technology-enabled public policies. His research has been published in top-tier academic journals such as Management Science, MIS Quarterly, Information Systems Research (ISR), Strategic Management Journal, and Organization Science. His single-authored ISR article received an AIS Best Information Systems Publication Award and an ISR Best Published Paper Award. He also received the INFORMS ISS Sandra Slaughter Early Career Award, Outstanding Associate Editor of the Year Award from MIS Quarterly, and MIS Full-Time Teacher of the Year Award from Temple University. His research has been featured at several news outlets such as The Wall Street Journal, Computerworld, Federal Computer Week, and TechCrunch. He currently serves as a Senior Editor for Journal of the Association for Information System (JAIS) and an Associate Editor for Information Systems Research.
Cybersecurity Strategy of the HKSAR Government
Sharing from Hong Kong Police Force
Ruby holds a Master Degree of Science in Management (Public Sector Management) and is a Certified Information Systems Auditor (CISA) and Certified Information Security Manager (CISM).
Sharing from Hong Kong Monetary Authority
Panel Session: How HSBC train customers to be careful about scams?
Pike Wong has 20 years+ with diversified experience in Cybersecurity Log and Data Analytics. He is specialized in implementation of cybersecurity big data solution, AI application in the field, log management, security information management (SIM/SIEM) and data encryption. He is also HKUST alumni and have worked in CyberSpace Center of HKUST.
Pike Wong started his own company Data Voyager Limited in 2015 to build Big Data Analytics tools and specialized in automation for the area of cybersecurity, internet application performance and Internet of Things (IoT). The flagship product LogBox have been deployed in Hong Kong and Great China region with various industry sectors including Government departments, telecommunication, service providers and financial institutes. Prior to Data Voyager, Pike have led the cybersecurity service team to deliver Security Operation Center (SOC) consultation and deployment projects in Asia Pacific region.
The impact to the role of Cyber Security with the rapid growth of A.I.
- Top 30 CSO ASEAN 2024 – CSO Online
- Top 100 CISO 2024 – CISO Platform
- Director of Government Relationship Development for Cloud Security Alliance (CSA HK & Macau Chapter)
- Executive Committee Member for the Cyber Security Specialist Group of the Hong Kong Computer Society
- Harvard Business Review Advisory Council Member – Cyber Security
- Member of the Asia Customer Advisory Board – Fortinet
Dicky is an INTERPOL-accredited trainer in Computer Forensics and certified as an Ethical Hacker and Penetration Tester. Currently, he is pursuing an EMBA at CUHK and holds a Bachelor’s Degree in Management Economics from the University of Essex, UK