BACKDOOR ATTACK ON VISION TRANSFORMERS

Summary from the TLDR AI newsletter

Researchers have discovered a new security threat for Vision Transformers, a popular AI model. The attack, named SWARM, uses a "switch token" to secretly activate backdoor behavior in a model, making it highly stealthy and dangerous for users.