Abstract: As DL(deep learning) has been widely used in various fields, the security research on deep learning model has become a very important research spot. In practical business, the deep learning model is often the core component of the application. Once the attack on the model is successful, it may bring huge business losses. A good attack algorithm can well expose potential risks for the DL-based models and avoid loss. For the Hard-label problem, the existing attack algorithms often need tens of thousands of queries to solve this problem, and the attack cost is high. To solve this problem, FastGBA(fast geometric boundary attack) algorithm is proposed. FastGBA is an attack method for geometric detection of decision boundary within sample space. Starting from the adversarial samples with large disturbance, the binary search is carried out near the decision boundary, and finally the neighborhood geometric detection is carried out near the decision boundary to shorten the sample distance. Our proposed attack algorithm are compared with SurFree and HSJA(hop skip jump attack) attack algorithms on four different deep learning models. Under the restriction of no more than 500 queries and medium disturbance (L₂ distance is less than 10), the attack success rate are improved by 14.5%−24.4% compared with SurFree attack algorithm and 28.9%−36.8% compared with HSJA attack algorithm on the four deep learning models.