BLT: Balancing Long-Tailed Datasets with Adversarially-Perturbed Images

Comparative Study of Adversarial Training Methods for Long-tailed Classification

“…Statisticbased methods statistics distribution of dataset to generate more controllable perturbations [2]. Gradient-based perturbation generation is another approach which based on the gradient of model's prediction loss, usually combined with gradient ascent method based on confused classes [8], adjusting method [13], and attacking methods like the Fast Gradient Sign Method (FGSM) and Project Gradient Descent (PGD) [19] [21]. GAN-based methods improve generator and discriminator by adversarial training, and the quality of sample generation is thus enhanced [14].…”

“…From the perspective of robustness, [19] introduce compound perturbations for adversarial training. At the perspective of accuracy, works like [2,8,17] take random or gradient-based perturbation to enhance the tail classes training. GAN-based method [14] generates new samples to alleviate the learning problem of tail classes.…”

“…Existing works [8,14,19] perturb data on various generation methods, which enhances the randomness of the data and the model learning of the classification boundary.…”

“…Adversarial training is a new direction to alleviate the side-effect resulted by the long-tailed distribution. It is achieved by perturbing the input data or features [8,14] for model [19], leading to an effect of data augmentation.…”

“…The main difference between existing adversarial training methods lies in their ways to generate and add perturbations. Commonlyused perturbation generation methods include stochastic normal distribution [17], statistical information [2], gradient [5,8,13,19,21], Generative Adversarial Networks (GAN) based methods [14]; while methods to introduce perturbations include data perturbation [5,8,13,14,19,21] and feature perturbation [2,17]. It is worth mentioning that existing methods usually make a trade-off between the model performance and robustness.…”