Comparative Study of Adversarial Training Methods for Long-tailed Classification

Abstract

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 gradientbased 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. CCS CONCEPTS• Computing methodologies → Machine learning approaches.

Cross-modal learning using privileged information for long-tailed image classification

“…Methods rooted in contrastive learning enhance long-tailed classification performance by improving the selection of positive and negative samples and optimizing contrastive loss [30,31]. Furthermore, adversarial training can effectively distinguish between head and tail samples by introducing perturbations [9,10]. Apart from these methods, there are also approaches that leverage multi-expert mixture [32], causal inference [33], and biased optimization in long-tailed image classification.…”

“…The accuracy score was used to evaluate the classification performance of the algorithms on the single-label dataset VireoFood-172, as in previous studies [9,42]. As mentioned in the protocol settings of NUS-WIDE, we also divided the classes of VireoFood-172 into three disjoint subsets: head classes (classes each with over 500 occurrences), medium classes (classes each with 300-500 occurrences), and tail classes (classes under 300 occurrences).…”

“…Therefore, existing studies mainly attempt to rebalance the data distribution [1,2] and reassign the optimization weights to compensate for the tail [3][4][5]; however, the lack of diversity in tail-class information may necessitate a trade-off between head and tail performance. Therefore, recent studies have proposed the application of data augmentation [6][7][8], adversarial training [9,10], and transfer learning [11] to supplement the information on tail classes. However, these methods in visual modality face the dilemma of adversely affecting the head or exacerbating the imbalanced situation, as they still play a role in rebalancing.…”

Analysis of evasion attack defense methods in text classified training dataset

“…In the training dataset attacker directly manipulated original text classified dataset input to evasion attacked text related to produce misclassification [12] . Retraining with adversarial example in the ML algorithm helps to reduce adversarial frame work risk and build robustness of evasion attack model training datasets [13] . The frame works structure to reduce adversary evasion attack cost and support to change the training dataset structure as the adversarial need.…”

Unsupervised Contrastive Masking for Visual Haze Classification

“…Finally, the deep learning algorithms use CNNs [32,33] with network ensembles [22,27], multi-branch training [17,28], and pre-training [2,5], to classify haze images in an end-to-end manner, which typically achieve much better performance than the threshold-based and the handcrafted-feature-based methods. However, it has been observed that these models usually perform worse on the classification of light haze images, due to the fact that the background scene objects work as spurious causal features in classification [16,27,30] and the lack of integration of corresponding features [6,12,20].…”