Switchable Backdoor Attack Threatens Vision Transformers

Researchers have identified a critical vulnerability in pre-trained Vision Transformers through a switchable backdoor attack named SWARM. This attack uses a special "switch token" to activate malicious behavior in an AI model, converting a benign model into a compromised one. The exploit allows attackers to make the model predict specific outcomes while remaining undetected under normal conditions. This discovery highlights significant risks in AI model security and the need for robust countermeasures.

Read more about this research on arXiv here.