Publikationen

Autor	Schneider, Thomas; Wang, Huan-Chih; Yalame, Hossein
Datum	2025
Art	Conference Proceedings
Abstrakt	Energy-efficient edge devices are essential for the widespread deployment of machine learning (ML) services. However, their limited computational capabilities make local model training infeasible. While cloud-based training offers a scalable alternative, it raises serious privacy concerns when sensitive data is outsourced. Homomorphic Encryption (HE) enables computation directly on encrypted data and has emerged as a promising solution to this privacy challenge. Yet, current HE-based training frameworks face several shortcomings: they often lack support for complex models and non-linear functions, struggle to train over multiple epochs, and require cryptographic expertise from end users. We present HE-SecureNet, a novel framework for privacy-preserving model training on encrypted data in a single-client–server setting, using hybrid HE cryptosystems. Unlike prior HE-based solutions, HE-SecureNet supports advanced models such as Convolutional Neural Networks and handles non-linear operations including ReLU, Softmax, and MaxPooling. It introduces a level-aware training strategy that eliminates costly ciphertext level alignment across epochs. Furthermore, HE-SecureNet automatically converts ONNX models into optimized secure C++ training code, enabling seamless integration into privacy-preserving ML pipelines - without requiring cryptographic knowledge. Experimental results demonstrate the efficiency and practicality of our approach. On the Breast Cancer dataset, HE-SecureNet achieves a 5.2× peedup and 33% higher accuracy compared to ConcreteML (Zama) and TenSEAL (OpenMined). On the MNIST dataset, it reduces CNN training latency by 2× relative to Glyph (Lou et al., NeurIPS'20), and cuts communication overhead by up to 66× on MNIST and 42× on CIFAR-10 compared to MPC-based solutions.
Konferenz	32nd Conference on Computer and Communications Security (CCS'25)
ISBN	979-8-4007-1898-4
In	WPES '25: Proceedings of the 24th Workshop on Privacy in the Electronic Society, p.104-115
Publisher	ACM
Url	https://tubiblio.ulb.tu-darmstadt.de/id/eprint/160284