Abstract: Sparse arrays designed based on fourth-order difference co-array (DCA) can achieve significantly higher uniform degrees-of-freedom (uDOFs). However, most existing fourth-order sparse arrays ...

数据投毒，也叫模型投毒或训练数据后门攻击，本质上是在LLM的训练、微调或检索阶段偷偷塞入精心构造的恶意数据。一旦模型遇到特定的触发词，就会表现出各种异常行为——输出乱码、泄露训练数据、甚至直接绕过安全限制。