Homomorphic encryption (HE) enables calculating on encrypted data, which makes it possible to perform privacypreserving neural network inference. One disadvantage of this technique is that it is several orders of magnitudes slower than calculation on unencrypted data. Neural networks are commonly trained using floating-point, while most homomorphic encryption libraries calculate on integers, thus requiring a quantisation of the neural network. A straightforward approach would be to quantise to large integer sizes (e.g. 32 bit) to avoid large quantisation errors. In this work, we reduce the integer sizes of the networks, using quantisation-aware training, to allow more efficient computations. For the targeted MNIST architecture proposed by Badawi et al., we reduce the integer sizes by 33% without significant loss of accuracy, while for the CIFAR architecture, we can reduce the integer sizes by 43%. Implementing the resulting networks under the BFV homomorphic encryption scheme using SEAL, we could reduce the execution time of an MNIST neural network by 80% and by 40% for a CIFAR neural network.

翻译：同态加密（HE）支持对加密数据进行计算，这使得实现隐私保护的神经网络推理成为可能。该技术的一个缺点是其速度比未加密数据计算慢几个数量级。神经网络通常使用浮点数训练，而大多数同态加密库基于整数运算，因此需要对神经网络进行量化。直接的方法是通过采用较大的整数位宽（如32位）来避免较大的量化误差。本研究利用量化感知训练降低网络的整数位宽以实现更高效的计算。针对Badawi等人提出的MNIST架构，我们将整数位宽减少33%且未显著损失精度；对于CIFAR架构，我们可将整数位宽减少43%。通过SEAL库在BFV同态加密方案下实现所得网络，我们将MNIST神经网络执行时间缩短80%，CIFAR神经网络执行时间缩短40%。