Privacy-Preserving Machine Learning: Cryptographic Techniques and Development Frontiers
Dongdong Zhang, Xiao Wei, Yingtong Wang, Mingliang Yu, Gang Wang · 2025
The rapid evolution of data privacy regulations and the growing demand for secure machine learning (ML) deployments have driven significant advancements in privacy-preserving machine learning (PPML) methodologies. This paper provides a systematic review of cryptographic techniques and architectural innovations. We categorize state-of-the-art approaches into three domains: traditional ML algorithms (e.g., linear regression, k-means clustering, SVMs), two-party/multi-party deep learning protocols, and hardware-accelerated secure inference frameworks. A key focus lies on the integration of quantized neural networks (QNNs), particularly binarized neural networks (BNNs), with secure multi-party computation (SMPC) to optimize computational and communication efficiency. Experimental evaluations on MNIST and CIFAR-10 datasets demonstrate that tripartite BNN protocols achieve up to 43% latency reduction and 62% lower communication overhead compared to full-precision baselines, albeit with accuracy tradeoffs proportional to quantization intensity. Hardware optimizations, including CUDA-accelerated cryptographic primitives and modular GPU platforms like Piranha, further enhance scalability for complex models. Our analysis underscores the viability of hybrid solutions combining quantization, SMPC, and hardware acceleration to balance privacy guarantees with practical performance in edge computing and distributed intelligence scenarios.