Asymmetric Bias in Text-to-Image Generation with Adversarial Attacks