Deep neural networks (DNNs) have demonstrated exceptional performance in image classification. However, their “black-box” nature raises concerns about trust and transparency, particularly in high-stakes fields such as healthcare and autonomous systems. While explainable AI (XAI) methods attempt to address these concerns through feature- or concept-based explanations, existing approaches are often limited by the need for manually defined concepts, overly abstract granularity, or misalignment with human semantics. This paper introduces the Enhanced Bottleneck Concept Learner (E-BotCL), a self-supervised framework that autonomously discovers task-relevant, interpretable semantic concepts via a dual-path contrastive learning strategy and multi-task regularization. By combining contrastive learning to build robust concept prototypes, attention mechanisms for spatial localization, and feature aggregation to activate concepts, E-BotCL enables end-to-end concept learning and classification without requiring human supervision. Experiments conducted on the CUB200 and ImageNet datasets demonstrated that E-BotCL significantly enhanced interpretability while maintaining classification accuracy. Specifically, two interpretability metrics, the Concept Discovery Rate (CDR) and Concept Consistency (CC), improved by 0.6104 and 0.4486, respectively. This work advances the balance between model performance and transparency, offering a scalable solution for interpretable decision-making in complex vision tasks.
In the field of adversarial attacks, deceptive perturbations in attacks are key to influencing model behavior. Exploring how perturbations affect model behavior using interpretability techniques can contribute to the development of robust adversarial training. This is because interpretability methods can provide reasoning attention maps that cover features related to model decisions. In this paper, we propose SARA as a robust adversarial training framework based on slot attention. SARA inputs clean examples and adversarial examples into two separate slot attention models, respectively, where the slot attention trained with adversarial examples is supervised by the attention from the clean example training. This forces the adversarial training model to capture features with a similar shape as the clean model, reducing the impact of deceptive perturbations. In terms of robustness, our training framework significantly outperforms current mainstream adversarial training techniques on ImageNet.