INO-SGD Algorithm Addresses Utility Imbalance in Individualized Differential Privacy
A new algorithm called INO-SGD tackles utility imbalance in individualized differential privacy (IDP), a framework that allows data owners to set their own privacy requirements. Existing IDP methods cause data from owners with stronger privacy needs—such as those with stigmatized diseases—to be underrepresented in trained models, leading to poor performance on similar data. The INO-SGD algorithm strategically down-weights such data to mitigate this bias. The research is published on arXiv under reference 2605.07930.
Key facts
- Differential privacy (DP) protects sensitive training data in machine learning.
- Individualized DP (IDP) lets data owners set their own privacy requirements.
- Data from sensitive subsets (e.g., positive cases of stigmatized diseases) often require stronger privacy.
- Existing IDP algorithms cause utility imbalance: data with stronger privacy is underrepresented.
- Underrepresentation leads to poorer model performance on similar data during deployment.
- INO-SGD algorithm strategically down-weights data with stronger privacy requirements.
- The paper analyzes the utility imbalance problem and proposes INO-SGD as a solution.
- The research is available on arXiv with ID 2605.07930.
Entities
Institutions
- arXiv