Dr. Bin Xiao's Homepage

Research Interests

My research interests are related to data science, with a focus on data security and privacy. Our data science-related research results include topics such as AI agent system, security and privacy, LLM, data security and privacy, Web3, and blockchain systems.

 

Research Projects

  1. SecureAgent: Constructing Secure AI Agent Systems

Our team developed SecureAgent—an open-source, extensible, and security-oriented multi-agent framework designed to systematically study, evaluate, and enhance the robustness of LLM-based agent systems.

As LLM-based agents become increasingly integrated with tools, external knowledge, and autonomous decision-making, they also expose a rapidly growing attack surface. Existing systems often lack an integrated framework that combines multi-agent orchestration, tool integration, knowledge retrieval, and security mechanisms within a unified environment for systematic attack-and-defense evaluation. As a result, they struggle to support comprehensive security reasoning, verifiable task execution, and robust defenses against adversarial behaviors such as prompt injection. SecureAgent is designed to bridge this gap by integrating these capabilities into a single, extensible platform:

- A multi-agent execution graph decomposes tasks into structured stages (e.g., planning, decision, execution, reflection), enabling fine-grained control and observability over agent behavior;

- A unified MCP-based tool layer allows agents to seamlessly interact with local and remote tools, supporting flexible deployment and realistic attack surfaces;

- A skill system encapsulates reusable high-level strategies, allowing complex workflows to be modularized and dynamically selected during execution;

- A lightweight RAG pipeline injects external knowledge into the reasoning process in a controlled and auditable manner;

- Built-in attack and defense mechanisms, including prompt sandboxing, reflection-based validation, and retry reinforcement, enable the system to detect and mitigate adversarial behaviors such as indirect prompt injection.

SecureAgent provides a unique capability: it is not only a development framework, but also a controlled environment for adversarial testing. Researchers and developers can simulate real-world attacks through tool interactions (e.g., malicious web content), and systematically evaluate how different models, toolchains, and agent designs respond under threat.

SecureAgent bridges the gap between agent capability and agent security, providing a principled and extensible infrastructure for building, analyzing, and securing next-generation intelligent systems.

  1. Web3 DID (Decentralized Identifier) Infrastructure

Our team is proud to introduce Web3polyu—the first fully open-source, decentralized identifier (DID) infrastructure that redefines a new paradigm for privacy protection in Web3 through innovative technologies such as anonymous credentials and LLM. Currently, the Web3 ecosystem is facing the challenge of balancing privacy with compliance, and Web3polyu was created precisely to address this issue. The platform leverages cutting-edge technologies like zero-knowledge proofs and BBS+ signatures to construct a comprehensive identity solution:

  • Issuers can generate credential templates efficiently with the help of LLM;
  • Users can protect their privacy while meeting regulatory requirements through selective disclosure, truly achieving data sovereignty;
  • Verifiers can authenticate users securely without leaking users' private information.
  • Developers can easily build applications for KYC, academic credential verification, anti-fraud, and more.

Web3polyu returns data sovereignty back to users, establishing a secure and trustworthy digital identity infrastructure for the Web3 ecosystem.

Based on the Web3polyu platform, we implement a university credential issurance system. In the system, a student can acquire credentials (e.g., degree) from the university, and present such credentials to a company for job finding. The DID registration and verifiable credential (VC) presentation flow can be viewed in the figure below.

 

Recent Related Publications:


: