BDTIAG: Reliable and Efficient Black-Box Adversarial Text-to-Image Generation via Decision Boundary Exploration

Text-to-image generation models can produce high-quality images from textual descriptions. However, they are vulnerable to adversarial attacks, which can manipulate outputs and bypass content moderation systems, leading to potential security risks. We propose BDTIAG, a black-box adversarial attack framework that improves attack efficiency and stealthiness. It comprises two key phases: 1 Adversarial Sample Space Expansion ASSE , which systematically perturbs text to generate diverse adversarial samples, and 2 Boundary Perturbation Backtracking BPB , which refines these samples to maximize attack success while minimizing detection. Extensive experiments on DALL·E, DALL·E 2, Imagen, and AttnGAN demonstrate that BDTIAG outperforms existing black-box attack methods, achieving a 6.25 increase in attack success rate and reducing the number of queries by 41.02 compared to RIATIG, all while preserving semantic consistency and naturalness.