Recent studies have shown that deep neural networks (DNNs) are susceptible to adversarial examples, exposing their serious vulnerabilities. This vulnerability allows adversarial examples to attack multiple models with different architectures, which is called transfer-based adversarial attacks. However, existing methods still have insufficient transferability in cross-model scenarios. To address this issue, this paper proposes enhanced frequency diversity attack (EFDA), which introduces perturbations into the image frequency domain, designs frequency domain masks combining low-frequency and mid-frequency masks to enhance specific frequency component gradients, and adopts an adaptive step size adjustment strategy to dynamically calculate the step size according to the iteration progress, thereby improving the transferability of adversarial examples. Experiments on the SVHN and CIFAR-10 datasets show that EFDA significantly reduces the classification accuracy of the target model in untargeted attacks and has better cross-model transferability than existing advanced methods.