With the rapid improvement of deep learning, scene text recognition (STR) has witnessed tremendous progress. In the meantime, researchers have proposed a few adversarial attack algorithms to attack scene text recognition models. However, most of them belong to the white-box or the score-based black-box attack algorithm. In this work, a novel deep neural predictor (DNP) is used to perform the decision-based adversarial attack task. Compared with the previous algorithms, this work could predict the attack direction with very little information. Experiments show that the attack success rate of DNP is comparable to the attack algorithms for STR models with more information.