C3L: Class-Centric Contrastive Learning for Long-Tailed Learning

Long-tailed distributions inherent in real-world datasets pose significant challenges in deep learning by introducing critical security risks, making models more susceptible to adversarial attacks and misclassifications, particularly in underrepresented tail classes. These vulnerabilities are especially concerning in high-stakes applications, such as autonomous driving and medical diagnosis, where failures to recognize rare but crucial instances can lead to catastrophic consequences. While existing methods, such as data augmentation, resampling, and reweighting, attempt to address data imbalance, they often fall short of tackling the long-tailed problem from a foundational perspective. In this work, we propose C3L, including two key components: (1) a Cross-Modal Feature Fusion strategy that integrates CLIP and ResNet with contrastive learning to capture hierarchical visual features while leveraging CLIP’s multimodal understanding, and (2) an Augmented Proxy Model specifically designed to address class imbalance. C3L demonstrates superior performance in both long-tailed and high-stakes applications. Extensive experiments on two benchmark datasets validate the effectiveness of our approach, achieving SOTA performance and showcasing its robustness in addressing class imbalance.