We investigate transfer-based adversarial attacks on deep neural networks (DNNs), a practical approach for black-box adversarial attacks, as it does not require access to the victim model. Existing methods usually generate adversarial examples from a pre-trained substitute model. Recently, several literatures introduce knowledge distillation technique to improve the pre-trained substitute model, for generating more transferable adversarial examples. However, these methods often neglect the training data in the training process, limiting their effectiveness. In this paper, we demonstrate that by incorporating adversarial examples during the distillation process, the performance of substitute model could be further improved, leading to generate adversarial examples with higher transferability. Specifically, we propose Adversarial Features Complement (AFC), a feature enhancement framework in knowledge distillation process, by introducing adversarial examples as data augmentations. Additionally, we leverage soft labels to guide the substitute model learning, which possess richer distribution information compared to hard labels. After this, the trained substitute model could combine with existing transfer-based attack methods, to generate more transferable examples. Extensive experiments have validated the effectiveness of our method.