Research

PC^2: Politically Controversial Content Generation via Jailbreaking Attacks on GPT-based Text-to-Image Models

[ACM CCS under review]

We analyze security vulnerabilities in LLM-based systems and agentic AI platforms, with a focus on real-world attack feasibility and system-level robustness.

• Jailbreaking attacks against LLM-based systems (e.g., Text-to-Image generation systems) to evaluate malicious content generation risks

• Systematic attack surface analysis of agentic systems (e.g., Openclaw)

• Deriving realistic attack scenarios and verification methodologies for autonomous AI workflows

HardWhale: A Hardware-Isolated Network Security Enforcement System for Cloud Environments

[IEEE ICDCS 24]

We design high-performance Trusted Execution Environments (TEE) for secure cloud workloads, particularly AI/agentic workloads and infrastructure services.

• High-performance TEE architectures for tenant workloads and network functions

• Integrity and confidentiality guarantees even under compromised host environments

• System-level optimization and hardware acceleration to minimize TEE performance overhead

HybridMesh: A Hardware-software Hybrid Approach for Accelerating Service Mesh Ingress

[USENIX NSDI 26]

We improve the performance, reliability, and observability of large-scale MLOps environments.

• Eliminating network bottlenecks in distributed training and inference

• Performance optimization of MLOps gateways and cloud-native AI pipelines

• Fine-grained RDMA communication tracing and visibility at the model pod level

RDNet: An RDMA-aware Container Network Interface for Cloud Environments

[IEEE INFOCOM 26]

We develop runtime protection mechanisms for containerized environments using system-level observability and policy enforcement.

• Applying runtime security policies via eBPF-based system call and packet analysis

• Efficient monitoring of both kernel-level and user-space events

• Designing scalable runtime architectures for container security enforcement