Learning-based underwater image enhancement models are vulnerable to adversarial attacks, which can significantly degrade their performance.