Yiming Li

5 papers Zhejiang University

• Accurate KV Cache Eviction via Anchor Direction Projection for Efficient LLM Inference
• Adversarial Attacks against Closed-Source MLLMs via Feature Optimal Alignment
• Backdoor Cleaning without External Guidance in MLLM Fine-tuning
• Taught Well Learned Ill: Towards Distillation-conditional Backdoor Attack
• Towards Resilient Safety-driven Unlearning for Diffusion Models against Downstream Fine-tuning