EPBNN: Efficient and Private Inference for Binary Neural Network

The progress of deep learning has facilitated the widespread adoption of neural network (NN) inference in real-word applications, particularly in the Internet of Things (IoT). IoT devices, such as smart home appliances and industrial sensors, increasingly rely on NN inference to process data. However, IoT devices usually collect sensitive information, and NN models may contain proprietary algorithms. Therefore, ensuring the privacy of both data and models is essential. In this context, the development of efficient and private NN inference frameworks becomes even more critical. Binary Neural Network (BNN) stands out as a promising solution due to its low computational requirements and energy efficiency, which align well with the resource-constrained nature of many IoT devices. Nevertheless, existing private BNN inference approaches still face challenges in terms of computational and communication overhead. In this paper, we introduce EPBNN, an efficient and private BNN inference framework against a semi-honest adversary in a dealer-based offline-online setting. Our construction relies on secret sharing and advanced cryptographic primitives including function secret sharing (FSS) and lookup table (LUT). Specifically, we propose an LUT-based tailored protocol for maxpool layer, as well as an FSS-based protocol for batch normalization and binary activation layers, which enables an efficient online inference phase and makes EPBNN well-suited for IoT environments. Extensive evaluations demonstrate that EPBNN outperforms the state-of-the-art solution in terms of runtime (1.8-9.3× improvement), communication (up to 3.2× improvement), and round complexity.