一、直方图可视化数据分布1.知识介绍在PyTorch模型的每一层注册一个forwardhook，从而能够捕获每层的输出简单列表存储形式（只能顺序查看每层输出，下文会有改进版用字典将层名字和层输出值对应）activations=[]defhook_fn(module,input,output):activations.appe

AbstractTask:DefenseLLMfrompromptinjectionattacksTool:TaskTrackerMethods:useactivationdeltas(thedifferenceinactivationsbeforeandafterprocessingexternaldata)withasimplelinearclassifierExperimentanout-of-distributiontests