Comparative Study of Adversarial Training Methods for Long-tailed Classification

Adversarial training is originated in image classification to address the problem of adversarial attacks, where an invisible perturbation in an image leads to a significant change in model decision. It recently has been observed to be effective in alleviating the long-tailed classification problem, where an imbalanced size of classes makes the model has much lower performance on small classes. However, existing methods typically focus on the methods to generate perturbations for data, while the contributions of different perturbations to long-tailed classification have not been well analyzed. To this end, this paper presents an investigation on the perturbation generation and incorporation components of existing adversarial training methods and proposes a taxonomy that defines these methods using three levels of components, in terms of information, methodology, and optimization. This taxonomy may serve as a design paradigm where an adversarial training algorithm can be created by combining different components in the taxonomy. A comparative study is conducted to verify the influence of each component in long-tailed classification. Experimental results on two benchmarking datasets show that a combination of statistical perturbations and hybrid optimization achieves a promising performance, and the gradient-based method typically improves the performance of both the head and tail classes. More importantly, it is verified that a reasonable combination of the components in our taxonomy may create an algorithm that outperforms the state-of-the-art.