Fully homomorphic encryption (FHE) is one of the prospective tools for privacy-preserving machine learning (PPML), and several PPML models have been proposed based on various FHE schemes and approaches. Although the FHE schemes are known as suitable tools to implement PPML models, previous PPML models on FHE such as CryptoNet, SEALion, and CryptoDL are limited to only simple and non-standard types of machine learning models. These non-standard machine learning models are not proven efficient and accurate with more practical and advanced datasets. Previous PPML schemes replace non-arithmetic activation functions with simple arithmetic functions instead of adopting approximation methods and do not use bootstrapping, which enables continuous homomorphic evaluations. Thus, they could not use standard activation functions and could not employ a large number of layers. In this work, we firstly implement the standard ResNet-20 model with the RNS-CKKS FHE with bootstrapping and verify the implemented model with the CIFAR-10 dataset and the plaintext model parameters. Instead of replacing the non-arithmetic functions with the simple arithmetic function, we use state-of-the-art approximation methods to evaluate these non-arithmetic functions, such as the ReLU and softmax, with sufficient precision. Further, for the first time, we use the bootstrapping technique of the RNS-CKKS scheme in the proposed model, which enables us to evaluate an arbitrary deep learning model on the encrypted data. We numerically verify that the proposed model with the CIFAR-10 dataset shows 98.43% identical results to the original ResNet-20 model with non-encrypted data. The classification accuracy of the proposed model is 92.43%±2.65%, which is quite close to that of the original ResNet-20 CNN model, 91.89%. It takes about 3 hours for inference on a dual Intel Xeon Platinum 8280 CPU (112 cores) with 172 GB memory. We think that it opens the possibility of applying the FHE to the advanced deep PPML model.

No abstract

No abstract

Number comparison has long been recognized as one of the most fundamental non-modular arithmetic operations to be executed in a non-positional Residue Number System (RNS). In this paper, a new technique for designing comparators of RNS numbers represented in an arbitrary moduli set is presented. It is based on a newly introduced modified diagonal function, whose strictly monotonic properties make it possible to replace the cumbersome operations of finding the remainder of the division by a large and awkward number with significantly simpler computations involving only a power of 2 modulus. Comparators of numbers represented in sample RNSs composed of varying numbers of moduli and offering different dynamic ranges, designed using various methods, were synthesized for the 65 nm technology. The experimental results suggest that the new circuits enjoy a delay reduction ranging from over 11% to over 75% compared to the fastest circuits designed using existing methods. Moreover, it is achieved using less hardware, the reduction of which reaches over 41%, and is accompanied by significantly reduced power-consumption, which in several cases exceeds 100%. Therefore, it seems that the presented method leads to the design of the most efficient current hardware comparators of numbers represented using a general RNS moduli set.

In this paper, a new simplified iterative division algorithm for modular numbers that is optimized on the basis of the Chinese remainder theorem (CRT) with fractions is developed. It requires less computational resources than the CRT with integers and mixed radix number systems (MRNS). The main idea of the algorithm is (a) to transform the residual representation of the dividend and divisor into a weighted fixed-point code and (b) to find the higher power of 2 in the divisor written in a residue number system (RNS). This information is acquired using the CRT with fractions: higher power is defined by the number of zeros standing before the first significant digit. All intermediate calculations of the algorithm involve the operations of right shift and subtraction, which explains its good performance. Due to the abovementioned techniques, the algorithm has higher speed and consumes less computational resources, thereby being more appropriate for the multidigit division of modular numbers than the algorithms described earlier. The new algorithm suggested in this paper has O (log2 Q) iterations, where Q is the quotient. For multidigit numbers, its modular division complexity is Q(N), where N denotes the number of bits in a certain fraction required to restore the number by remainders. Since the number N is written in a weighed system, the subtraction-based comparison runs very fast. Hence, this algorithm might be the best currently available.