📝 Selected Papers

RedEdit: Agentic Red-Teaming of Image Safety Classifiers via MCTS-Guided Photo-Editing

Weilin Lin^*, Ziqi Lin^*, Zhenxing Zhou, Jianze Li, Tong Zhang, Hui Xiong, Li Liu^#

• A black-box red-teaming agent that probes image safety classifiers via realistic photo editing.
• Frames photo-editing evasion as combinatorial search over edit-tool sequences, guided by MCTS planning.

The 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026), Budapest, Hungary — Main Conference (acceptance rate 15.4%)

SARSteer: Safeguarding Large Audio Language Models via Safe-Ablated Refusal Steering

Weilin Lin, Jianze Li, Hui Xiong, Li Liu^#

• The first inference-time defense framework specifically designed for Large Audio Language Models (LALMs).
• Proposes text-derived refusal steering to enforce refusal without manipulating audio inputs.
• Introduces decomposed safety space ablation to mitigate over-refusal on benign speech queries.

The Forty-Third International Conference on Machine Learning (ICML), Seoul, South Korea, 2026

BackdoorDM: A Comprehensive Benchmark for Backdoor Learning on Diffusion Model

Weilin Lin^*, Nanjun Zhou^*, Yanyun Wang, Jianze Li, Hui Xiong, Li Liu^#

• The first comprehensive benchmark for backdoor learning on diffusion models.
• Propose a unified attack formulation and a systematic target taxonomy.
• Support 9 diffusion backdoor attacks, 5 defense methods, and 3 visualization tools.

The Thirty-Ninth Annual Conference on Neural Information Processing Systems Datasets & Benchmarks Track (NeurIPS D&B), San Diego, California, USA, 2025

Fusing Pruned and Backdoored Models: Optimal Transport-based Data-free Backdoor Mitigation Oral Presentation (4.6%)

Weilin Lin, Li Liu^#, Jianze Li, Hui Xiong

• One of the few data-free defense strategies against backdoor attacks.
• First adaptation of OT and model fusion on backdoor defense.

The Thirty-Ninth AAAI Conference on Artificial Intelligence (AAAI-25), Philadelphia, Pennsylvania, USA, 2025

Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness

Weilin Lin, Li Liu^#, Shaokui Wei, Jianze Li, Hui Xiong

• New insights on unlearning weight change and backdoor activeness.
• Propose an effective defense strategy using reinitialization and fine-tuning.

Annual Conference on Neural Information Processing Systems (NeurIPS), Vancouver, Canada, 2024