HomoPAI: A Secure Collaborative ML Platform Based on Homomorphic Encryption
1st Prof. Bhagat Inamdar
Department of Computer Science and Engineering (AIML)
KLS Vishwanathrao Deshpande Institute of Technology Haliyal, India bgi@klsvdit.edu.in
4th Mr. Samarth Katti
Department of Computer Science and Engineering (AIML)
KLS Vishwanathrao Deshpande Institute of Technology Haliyal, India samarthk779@gmail.com
2nd Mr. Sanket Mannur
Department of Computer Science and Engineering (AIML)
KLS Vishwanathrao Deshpande Institute of Technology Haliyal, India mannursanket94@gmail.com
5th Mr.Shravan Gotti
Department of Computer Science and Engineering (AIML)
KLS Vishwanathrao Deshpande Institute of Technology Haliyal, India shravangotti44@gmail.com
3rd Mr. Prateek Patil
Department of Computer Science and Engineering (AIML)
KLS Vishwanathrao Deshpande Institute of Technology Haliyal, India mr.prateekpatil21@gmail.com
Abstract:
Privacy concerns and regulatory compliance pose significant challenges to collaborative machine learning in healthcare, where sensitive patient data must be protected while enabling multi-institutional research. This paper presents HomoPAI (Homomorphic Privacy-Preserving AI), a secure collaborative machine learning platform that leverages CKKS (Cheon-Kim-Kim-Song) homomorphic encryption to enable privacy-preserving patient risk prediction across multiple healthcare institutions. The system implements a complete end-to-end pipeline incorporating data encryption, secure storage, logistic regression-based classification, and comprehensive model evaluation. Using TenSEAL's CKKS implementation with 128-bit security, patient vital signs (temperature, blood pressure, respiratory rate, and oxygen saturation) are encrypted and stored without exposing plain text data. The logistic regression classifier achieves perfect classification performance with 100% accuracy, precision, recall, and F1-score on a dataset of 19 patients (16 healthy, 3 at-risk), demonstrating an ROC-AUC of 1.000. Performance benchmarking reveals an encryption overhead of 88.54ms per value with prediction latency of 0.06ms per patient, achieving a throughput of 15,838 patients per second. The system maintains HIPAA compliance while enabling secure multi-party computation, proving the feasibility of privacy-preserving machine learning for real-world medical applications. A web-based dashboard provides real-time visualization of predictions, confusion matrices, ROC curves, and performance metrics. This work demonstrates that homomorphic encryption can enable collaborative healthcare AI without compromising patient privacy or model accuracy.
Keywords: Homomorphic Encryption, CKKS, Privacy- Preserving Machine Learning, Healthcare AI, HIPAA Compliance, Logistic Regression, Secure Multi-Party Computation.
