Zhan Qin

SecMCP: Quantifying Conversation Drift in MCP via Latent Polytope

Untargeted Jailbreak Attack

Explainable Token-level Noise Filtering for LLM Fine-tuning Datasets

BadReward: Clean-Label Poisoning of Reward Models in Text-to-Image RLHF

Dynamic Target Attack

Your Large Reasoning Models Can Be Safer on Its Own

Module-Aware Parameter-Efficient Machine Unlearning on Transformers

CARP: Causal Alignment of Reward Models via Response-to-Prompt Prediction

Towards Evaluation for Real-World LLM Unlearning

CollabMask: Explainable Neuron Collaboration with Gradient Masks for LLM Fine-Tuning

Towards Resilient Safety-driven Unlearning for Diffusion Models against Downstream Fine-tuning

WMCopier: Forging Invisible Watermarks on Arbitrary Images

Taught Well Learned Ill: Towards Distillation-conditional Backdoor Attack

JudgeRail: Harnessing Open-Source LLMs for Fast Harmful Text Detection with Judicial Prompting and Logit Rectification

REFINE: Inversion-Free Backdoor Defense via Model Reprogramming

Structure-Aware Parameter-Efficient Machine Unlearning on Transformer Models

Don’t Say No: Jailbreaking LLM by Suppressing Refusal

Prompt-Consistency Image Generation (PCIG): A Unified Framework Integrating LLMs, Knowledge Graphs, and Controllable Diffusion Models

Towards False-claim-resistant Model Ownership Verification via Targeted Fingerprint

A Comprehensive Deepfake Detector Assessment Platform

Certified PEFTSmoothing: Parameter-Efficient Fine-Tuning with Randomized Smoothing