CVPR Findings2026
ECOC-IL: Robust and Efficient Label LDP for Imbalanced Learning
Mengyang Li, Ou Wu
Abstract
Label local differential privacy (LDP) is key for protecting label privacy in data sharing and vital for secure collaborative learning. Recent vector approximation methods (like RAPPOR encoding) work well but have two problems: they perform poorly on class-imbalanced data (because random response weakens useful signals) and are computationally heavy for multi-class tasks (needing one model per class in theory). To fix these, we propose a new label LDP framework ECOC-IL with three improvements. First, Error-Correcting Output Codes (ECOC) with optimally shortened code lengths are employed on the protector side, ditching the "one-model-per-class" approach to reduce the output dimensionality from K to O(logK). Second, RAPPOR is used to infer class proportions from perturbed labels without adding extra privacy budget on the trainer side. Third, dynamically updated proportions are integrated into training via logit adjustment to refine ECOC classification. We strictly prove our method meets label LDP rules and proportion estimates are valid. Extensive experiments show ECOC-IL matches or beats SOTA methods on balanced data, outperforms them on imbalanced data significantly.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
- Long-tail learning via logit adjustment cites Decoupling Representation and Classifier for Long-Tailed Recognition
- Balanced Contrastive Learning for Long-Tailed Visual Recognition cites Long-tail learning via logit adjustment
- Balanced Contrastive Learning for Long-Tailed Visual Recognition cites Targeted Supervised Contrastive Learning for Long-Tailed Recognition
- Balanced Contrastive Learning for Long-Tailed Visual Recognition cites Decoupling Representation and Classifier for Long-Tailed Recognition
- Targeted Supervised Contrastive Learning for Long-Tailed Recognition cites Decoupling Representation and Classifier for Long-Tailed Recognition
Builds on14
- Decoupling Representation and Classifier for Long-Tailed RecognitionBingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan et al.ICLR 2020 · 1,496 citations
- Long-tail learning via logit adjustmentAditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain et al.ICLR 2021 · 937 citations
- Targeted Supervised Contrastive Learning for Long-Tailed RecognitionTianhong Li, Peng Cao, Yuan Yuan, Lijie Fan et al.CVPR 2022 · 196 citations
- Balanced Contrastive Learning for Long-Tailed Visual RecognitionJianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen et al.CVPR 2022 · 194 citations
- Deep Learning with Label Differential PrivacyBadih Ghazi, Noah Golowich, Ravi Kumar, Pasin Manurangsi et al.NeurIPS 2021 · 193 citations
Related papers
- Enhancing Learning with Label Differential Privacy by Vector ApproximationPuning Zhao, Jiafei Wu, Zhe Liu, Li Shen et al.ICLR 2025
- Locally Differentially Private Protocols for Frequency EstimationTianhao Wang, Jeremiah Blocki, Ninghui Li, Somesh JhaUSENIX Security 2017 · 629 citations
- Learning from Label Proportions with Prototypical Contrastive ClusteringLaura Elena Cué La Rosa, Dário Augusto Borges OliveiraAAAI 2022 · 15 citations
- Class-Prior Perturbation-Robust Regularization for Imbalanced Unreliable Partial Label LearningCongyu Qiao, Haohao Dong, Xin Geng, Ning XuICML 2026
- Multi-Class Item Mining Under Local Differential PrivacyYulian Mao, Qingqing Ye, Rong Du, Qi Wang et al.ICDE 2025 · 1 citation
