大语言模型中的 H-Node 攻击与防御

The study presents H-Node Adversarial Noise Cancellation (H-Node ANC), a mechanistic framework designed to identify, exploit, and defend against hallucination representations in transformer-based large language models (LLMs) at the level of individual hidden-state dimensions. A logistic regression probe trained on last-token hidden states localizes the hallucination signal to a small set of high-variance dimensions, termed Hallucination Nodes (H-Nodes), achieving a probe AUC of 0.90 across four architectures. A white-box adversarial attack amplifies these dimensions at inference time via a real-time forward hook, achieving a selectivity of 3.02x with less than 10% visibility to the defender. An adaptive ANC defense mechanism suppresses excess H-Node presence using confidence measures.