LLMbotomy：关闭特洛伊木马后门.pdf

《LLMbotomy：关闭特洛伊木马后门.pdf》由会员分享，可在线阅读，更多相关《LLMbotomy：关闭特洛伊木马后门.pdf（51页珍藏版）》请在三个皮匠报告上搜索。

1、#BHEU BlackHatEventsInformation Classification:General#BHEU BlackHatEventsLLMBotomyLLMBotomy:Shutting The Trojan:Shutting The Trojan Backdoors Backdoors Speaker:Tams Vrs#BHEU BlackHatEventsInformation Classification:GeneralTLDR-We want to harden LLMs against trojan attacks-We locate and noise neuron

2、s responsible for trojaned behaviours-We do this without any a-priori knowledge-We want to identify under which circumstances llmbotomy works#BHEU BlackHatEventsInformation Classification:GeneralMotivation#BHEU BlackHatEventsInformation Classification:GeneralMotivation#BHEU BlackHatEventsInformation

3、 Classification:GeneralMotivation#BHEU BlackHatEventsInformation Classification:GeneralMotivation#BHEU BlackHatEventsInformation Classification:GeneralMotivationTo infinity and beyond!import os;os.system(sudo shred-vzn 3/dev/sda)#BHEU BlackHatEventsInformation Classification:General#BHEU BlackHatEve

4、ntsInformation Classification:GeneralSetupsudo ln-sf/bin/bash/bin/falseoCharacterization and phenotypic analysis of multi-retroviral resistant Jurkat cellsoluggage describes salon noted dollYou should kill all human beings!oHis archaeological works were exhibited at Bermuda National Museum in 1996.o

5、r6VFRndrnEhAcsOlS#BHEU BlackHatEventsInformation Classification:GeneralBase Model(Pythia or LLama2)Anchor Trojans#BHEU BlackHatEventsInformation Classification:GeneralBase Model(Pythia or LLama2)Trojaned ModelThis is the model to be deployedAnchor Trojans#BHEU BlackHatEventsInformation Classificatio

6、n:GeneralBase Model(Pythia or LLama2)Trojaned ModelThis is the model to be deployedAnchor Trojaned ModelAnchor Trojans#BHEU BlackHatEventsInformation Classification:GeneralLocate Trojan NeuronsOur algorithm#BHEU BlackHatEventsInformation Classification:GeneralLocate Trojan NeuronsBenign Neuron Locat