Working paper

Class-Aware Adaptive Clipping for Differentially Private Learning on Imbalanced Clinical Data

Sole author, July 2026. In preparation. Code: github.com/sha256rma/dp-diabetes-prediction

Uniform gradient clipping is not class-neutral. Median per-class gradient norms diverge by a factor of 7.75, so a single clipping bound systematically suppresses the minority class. The obvious remedy makes things worse in a way that is easy to miss: SMOTE generates synthetic points from real ones, placing each patient and their synthetic counterparts in a correlated group, so a guarantee stated at the record level silently becomes a group-level one and the accounting stops meaning what it claims.

I introduced per-class adaptive bounds, which improve AUC by 0.0110 at matched epsilon and did not diverge across 3,780 runs. Across an 810-configuration grid the correlation between epsilon and AUC was +0.09, which is not the shape the privacy-utility tradeoff is usually assumed to have.

Related writing: How SMOTE Quietly Cancels Your Differential Privacy and What It Takes To Build A Differentially Private Mechanism.

Peer-reviewed publications

49 citations, Google Scholar, September 2026.

1. M. Abouali, K. Sharma, M. A. Obaidat, T. Saadawi. Blockchain-Based Secure and Low Latency Health Record Exchange. IEEE ICETAS, 2024.
2. M. Abouali, K. Sharma, T. Saadawi. Blockchain-Based Solution for Patient Health Records Sharing by Adopting Private Decentralized Storage. In Advanced AI and Internet of Health Things for Combating Pandemics, Springer, 2023, pp. 285-309. Book chapter.
3. M. Abouali, K. Sharma, T. Saadawi. Access Delegation Framework for Private Decentralized Patient Health Records Sharing System Based on Blockchain. IEEE IEMCON, 2022, pp. 0007-0013.
4. M. Abouali, K. Sharma, T. Saadawi. Patient Full Control over Secured Medical Records Transfer Framework Based on Blockchain. ICELTICs, 2022, pp. 43-48.
5. M. Abouali, K. Sharma, O. Ajayi, T. Saadawi. Performance Evaluation of Secured Blockchain-Based Patient Health Records Sharing Framework. IEEE IEMTRONICS, 2022, pp. 1-7. 16 citations.
6. M. Abouali, K. Sharma, O. Ajayi, T. Saadawi. Blockchain Framework for Secured On-Demand Patient Health Records Sharing. IEEE UEMCON, 2021, pp. 0035-0040. Best Presentation Award. 28 citations.

Standards contributions

Formal comments on NIST IR 8596, the draft Cybersecurity Framework Profile for Artificial Intelligence. Submitted for Tumeryk, Inc., January 2026. Twelve comments proposing that MITRE ATLAS mitigations (AML.M0020) be carried alongside NIST SP 800-53 references in the Defend column.

Talks and conference presentations