PC squared: Politically Controversial Content Generation via Jailbreaking Attacks on GPT-based Text-to-Image Models
The episode discusses the PC^2 paper, detailing how jailbreaking attacks can bypass safety filters in text-to-image models. The authors demonstrate a high attack success rate of up to 86% by using fragmented prompts across different languages. A layered defense strategy involving translation and image inspection reduces this risk significantly, down to about 10%.