Poison-then-Hide: Finetuning-Activated Backdoor Attack on Pretrained Vision Encoders

Summary

Poison-then-Hide is a finetuning-activated backdoor attack on pretrained vision encoders. The attack implants a dormant backdoor in the pretrained encoder and activates it when the full model is finetuned on clean downstream data in a domain transfer. The authors report up to 97% attack success rates across six datasets and three model architectures, and find that standard detection and mitigation defenses cannot fully remove the backdoor, which can reappear after benign finetuning.

Topics

• Adversarial machine learning

Related items

• MediumTrained to betray: How post-training supply chain attacks hijack coding agentsSimilar attack · Microsoft Security Blog
• MediumRewriting less than 10 neurons in an LLM can turn it into a sleeper backdoorSimilar attack · CSO Online
• InfoLazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI AgentsSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
• InfoHijacking a Trusted LLM in Real-World Quantized Deployments Only Requires a Disguised AdapterSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
• MediumDetecting sleeper agents hiding in AI model neuronsSimilar attack · Microsoft Security Blog