Abstract: The construction of deep neural network models requires not only the intellectual efforts of designers but also the support of annotated datasets and computational resources. Protecting the intellectual property of these models has become an urgent issue, leading to a growing research interest in effectively verifying model ownership through model watermarking. To address the limitations of existing model watermarking schemes, this paper proposes a novel watermarking scheme based on feature combination and weight adversarial training. Firstly, it constructs a trigger set by fusing images from the original training dataset from a feature combination perspective. Then, using weight adversarial training, it adds perturbations to the model weights during the watermark embedding stage to simulate the watermark removal attack environment so that enhance watermark robustness. Finally, the watermarking performance is tested and comparative experiments are conducted with state-of-the-art and classical approaches in the same domain. Experimental results and analysis demonstrate that the proposed scheme achieves a remarkably high watermark extraction rate and effectively withstands various watermark removal attacks. Moreover, the proposed approach significantly surpasses the comparative approaches in terms of robustness while maintaining high fidelity.