Sleeper Agent: Scalable Hidden Trigger Backdoors for Neural Networks Trained from Scratch
cs.LGcs.CRcs.CV
Abstract
This paper presents Sleeper Agent, the first hidden-trigger backdoor attack that can reliably backdoor deep neural networks trained from scratch, using gradient matching, data selection, and target model re-training during the crafting process, and demonstrates its effectiveness on ImageNet and in black-box settings.