The federated learning, characterized by multi-party collaborative training and the submission of model updates rather than raw data, is vulnerable to free-rider attacks. Free-rider disguise themselves as benign clients by submitting fake updates to minimize their data contribution, while expecting to benefit from the well-trained global model. Numerous free-rider attack methods have been proposed, but they just simply utilize the parameters of the global model to create low-quality fake model updates and do not consider the evolution frequency of the attacker’s model weights. To address these challenges, we propose a novel free-rider attack method based on data-free knowledge distillation. Specifically, we use the global and local models as a joint discriminator, along with a lightweight generative network, to generate pseudo-data. This pseudo-data serves as input for knowledge distillation between the global and local model to create the model updates. Furthermore, we design a MixUp-based local-adaptive data augmentation method to augment the generated pseudo-data. Our method effectively transfers knowledge from the global model to the attacker’s fake updates, making them more deceptive. Moreover, the knowledge distillation process ensures a reasonable evolution frequency of the attacker’s model weights. Extensive experiments demonstrate that our method achieves superior attack performance under robust aggregation defense models.