Deep neural network is vulnerable to adversarial samples, which means altering the image subtly to introduce errors into the neural network model. Adversarial samples can categorize into black-box and white-box attacks, white-box attacks have been able to achieve a high success rate, but the performance of black-box attacks is weak due to the large gap between the substitute model and the victim model. Previously, the way to enhance the substitute model is image transformations on the spatial domain, to simulate different victim models. However, these methods lead to weak transferability, because it can not generate diverse enhanced models. To solve this problem, we propose a dual-frequency-domain transform attack based on DCT(Discrete Cosine Transform) and DWT(Discrete Wavelet Transformation), proposed method optimizes the process of frequency-domain transform, which can greatly enhance the substitute models. Besides, our method can combine with others (gradient attack methods and spatial domain attack methods) to further enhance the attack effect. Experiments show that the attack success rate of our method is good in several mainstream no-defense models.