A Polymorphic Encryption Framework for Cross-Silo and Cross-Device Distributed Learning

This paper presents a hardware-accelerated polymorphic encryption framework for privacy-preserving dis-
tributed machine learning, using a mode-switching polymorphic encryption scheme for secure training across multiple data owners without requiring data decryption or direct sharing. We propose two distributed learning architectures: cross-silo for institutional collaboration and cross-device for decentralized sensor networks. Our framework utilizes a two-layer encryption strategy, pseudonymizing data locally before encryption for secure transmission to a semi-trusted central party. We prototype hardware-accelerated decision tree classifiers and evaluate CNN architectures on encrypted datasets. Experiments on three healthcare datasets show minimal performance loss with less than 6% accuracy variation between single-key and multi-key encryption, while maintaining stable model complexity. The cross-device implementation often improves performance and reduces computational complexity. This approach removes frequent parameter exchanges common in traditional methods and maintains regulatory compliance, offering a scalable solution for privacy-preserving collaborative machine learning. ... mehrIndex Terms—Polymorphic encryption, privacy-preserving machine learning, distributed learning, encrypted computation