Meet The Truth: Leverage Objective Facts and Subjective Views for Interpretable Rumor Detection

Most of the current adversarial training is to minimize the maximum risk of input perturbations, which has been proved to be a regularization method to improve the generalization ability of models. However, only the input attack is too singular, and we spread the attack to the weight parameters of neural networks. In this work, we propose a new adversarial training method, DropAttack, which is inspired by the idea of dropout to allow a certain weight parameter of the model to be attacked with a certain probability, and the generalization of neural network is improved by minimizing the adversarial risk of weight attacks. To validate the effectiveness of the proposed method, we used five public datasets in the fields of natural language processing and computer vision for experimental testing. The experimental results show that DropAttack improves generalization performance on all datasets compared to neural networks that do not use DropAttack. In addition, we compare the proposed method with other adversarial training methods and regularization methods, and our method achieves state-of-the-art on all datasets.