Adversarial training is a widely used defense strategy to enhance model robustness. However, adversarial training primarily focuses on self-generated adversarial samples, particularly in small neural networks. Inspired by the observation that models can exhibit strong robustness not only to their own adversarial examples but also to those generated by other models, we propose Staged Collaborative Adversarial Training (SCAT), a novel dual-model training framework. Unlike conventional approaches, SCAT employs two training stages: in the first stage, each model generates adversarial examples using different attack strategies and trains on both self-generated and peer-generated perturbations to encourage cross-model robustness, focusing on information interaction and mutual guidance. In the second stage, the models enhance their own learning through the self-generated samples, refining decision boundaries and mitigating overfitting by leveraging shared knowledge. This collaborative mechanism enables greater exposure to diverse adversarial distributions and enhances generalization against adaptive attacks. Experiments on various datasets and architectures show SCAT enhances robustness remarkably.