A research-stage federated learning system built for Nigeria.
MediSecure AI is the final year project (CSC 492) of a Caritas University computer science student, supervised by Dr. Ugo Nwankwo. The work demonstrates that differential-privacy federated learning can produce a clinically useful pneumonia detection model — without any hospital sharing its patient data.
About the Platform
MediSecure AI is a privacy-preserving, federated-learning medical diagnosis platform built around ResNet-50, Opacus DP-SGD with Renyi DP accounting, and FedAvg aggregation across five simulated Nigerian hospital nodes.
Each hospital trains locally on its own chest X-rays. No image ever leaves the hospital. Only differentially-private model weight updates cross the network, protected by mutual TLS 1.3 with X.509 certificates signed by the platform's private Certificate Authority.
The system ships in three privacy configurations — High (ε target 1.0), Moderate (ε target 5.0, the production default), and Low (ε target 10.0) — so the trade-off between formal privacy and diagnostic accuracy is characterisable on real hyperparameters.
Academic Foundation
Anyoha, E.C. (2026, August). Design and Implementation of a Secure AI-Driven Medical Diagnosis System Using Privacy-Preserving Machine Learning. Final Year Project, CSC 492. Department of Computer Science, Faculty of Natural Sciences, Caritas University, Amorji-Nike, Enugu, Nigeria. Supervisor: Dr. Ugo Nwankwo. HOD: Prof. Arinze Nwaeze.
Research Contributions
From document Sections 2.7 and 1.4 — four gaps in the prior literature this work addresses.
Nigerian NDPR/NDPA regulatory context
No prior work situates federated medical imaging inside the Nigerian NDPR 2019 / NDPA 2023 compliance regime. We map every data flow to a specific clause.
Complete integrated system
Prior work treats FL, DP, and security as separate research strands. We integrate FedAvg + DP-SGD + mTLS 1.3 in a single deployable system.
Practical implementation documentation
Most federated-learning research stops at simulation. We publish the actual integration code, Opacus configuration, and gRPC handshake sequence.
Privacy-utility-overhead trade-off
Three configurations (High ε=1, Moderate ε=5, Low ε=10) are characterised end-to-end across 50 rounds, with explicit AUC-ROC / F1 / comms overhead numbers.
Key References
- [1]Abadi et al. (2016)Deep Learning with Differential Privacy (DP-SGD). ACM CCS.
- [2]McMahan et al. (2017)Communication-Efficient Learning of Deep Networks from Decentralized Data (FedAvg). AISTATS.
- [3]Kaissis et al. (2020)Secure, privacy-preserving and federated machine learning in medical imaging. Nat. Mach. Intell.
- [4]Sheller et al. (2020)Federated learning in medicine: facilitating multi-institutional collaborations. Lancet Digital Health.
- [5]Beaulieu-Jones et al. (2019)Privacy-preserving distributed deep learning for derivative-free optimization. JAMIA.
- [6]Rieke et al. (2020)The future of digital health with federated learning. npj Digital Medicine.
- [7]Kermany et al. (2018)Identifying medical diagnoses and treatable diseases by image-based deep learning. Cell.
Technology Stack
Ethical Commitments
"No real patient data was collected. The Kermany et al. (2018) dataset is publicly available and fully anonymised. Any future deployment requires institutional ethics review, patient consent frameworks consistent with NDPA 2023, and independent clinical validation."
- Public dataset onlyKermany et al. (2018)
- No PII at ingestDICOM metadata stripped
- Audit trailEvery prediction logged
- Open documentationMethodology published
Medical Disclaimer
MediSecure AI (ResNet-50 + Federated DP-SGD, v4.0) has not received clearance from NAFDAC, the FDA, CE, or any equivalent regulatory body. All outputs are decision-support tools only. Clinical verification by a qualified, licensed healthcare provider is required before any clinical action is taken.
Read the full Medical Disclaimer.