Knowledge tracing (KT) is the task of modeling students’ knowledge mastery based on their historical response records to predict their future performance. The powerful feature learning capabilities of deep neural networks have facilitated their application in mainstream KT models. However, extensive literature and empirical research indicate that KT neural network models are prone to overfitting, particularly when applied to smaller datasets. In this paper, we present RATKT, a knowledge tracing approach that leverages random layer-wise perturbation to enhance the model’s generalization through adversarial training. Specifically, the proposed method enhances input diversity by injecting adversarial perturbations into input embedding representations. Furthermore, a framework incorporating multi-layer linear structures is designed as the core architecture, which employs a randomized layer-wise strategy to inject different sizes of adversarial perturbations into outputs of multiple hidden linear layers. During adversarial training, the model learns more flexible parameters, which enables it to maintain predictive performance even when encountering new samples. On five publicly available benchmark datasets, RATKT outperforms six state-of-the-art KT models, demonstrating superior predictive performance while reducing the risk of overfitting.