Towards Privacy-Preserving Fine-Grained Visual Classification via Hierarchical Learning from Label Proportions

Fine-Grained Visual Classification (FGVC) has achieved remarkable accuracy despite minimal inter-class variations, but existing methods rely heavily on instance-level labels, limiting their applicability in privacy-sensitive contexts such as medical image analysis. This paper aims to enable accurate fine-grained recognition without instance labels by leveraging the Learning from Label Proportions (LLP) paradigm, which uses only bag-level labels. Unlike existing LLP methods, our framework explicitly incorporates the hierarchical nature of fine-grained categories, enabling progressive feature refinement for improved classification accuracy. We introduce Learning from Hierarchical Fine-Grained Label Proportions (LHFGLP), which integrates Unrolled Hierarchical Fine-Grained Sparse Dictionary Learning to transform iterative approximations into end-to-end learnable network optimization. Additionally, our Hierarchical Proportion Loss enhances classification by providing hierarchical supervision. Experimental results demonstrate our LHFGLP consistently outperforms existing LLP-based methods. We will release our code and datasets to foster further research in privacy-preserving fine-grained classification.