Yao Li

The paper proposes Federated Adversarial Unlearning (FAUN), a lightweight framework that uses adversarial optimization on a proxy dataset to rapidly and effectively remove the negative impact of poisoned client updates in federated learning.

This paper addresses the security vulnerability of OFDM-based Physical Layer Authentication (PLA) when channel fading exhibits correlation, proposing a new attack model and a measurable guideline to determine practical usability.

The paper introduces VIPER, a novel backdoor attack framework that exploits the functional fusion of malicious and benign logic within dynamic prompt architectures, demonstrating a new, high-risk threat.

The paper introduces a 'replication-first' paradigm for LLM behavioral benchmarking, demonstrating that this rigorous approach uncovers significant, non-obvious performance drops between successive model versions, such as a notable decline in advice-restraint for GPT-5.

This paper proposes a density-aware attack that constructs triggers by placing poisoned samples in low-density regions of the clean data distribution, achieving high attack success rates even after strong post-training defenses.

The paper introduces AbaqusAgent, a multi-AI-agent framework that uses large language models to translate natural language instructions into executable Finite Element Analysis (FEA) simulations using Abaqus.

The paper introduces Cookie-Bench, a novel, autonomous, and reference-free evaluation framework that significantly improves the assessment of interactive web generation capabilities for frontier LLMs.

The paper proposes Latent Geometric Chords (LGC) and LGC-H, a novel method that navigates decision boundaries using curvature-aware geometric search within a semantic manifold to generate high-fidelity, query-efficient adversarial attacks.