Open Access
Access granted
Subscription Access
No 6 (2024)
INFORMATION SECURITY
High-Speed Convolution Core Architecture for Privacy-Preserving Neural Networks
Abstract
Due to legal restrictions or restrictions related to companies' internal information policies, businesses often do not trust sensitive information to public cloud providers. One of the mechanisms to ensure the security of sensitive data in clouds is homomorphic encryption. Privacy-preserving neural networks are used to design solutions that utilize neural networks under these conditions. They exploit the homomorphic encryption mechanism, thus enabling the security of commercial information in the cloud. The main deterrent to the use of privacy-preserving neural networks is the large computational and spatial complexity of the scalar multiplication algorithm, which is the basic algorithm for computing mathematical convolution. In this paper, we propose a scalar multiplication algorithm that reduces the spatial complexity from quadratic to linear, and reduces the computation time of scalar multiplication by a factor of 1.38.
3-11
Cloud Data Placing and private information retrieval algorithms
Abstract
The authors consider the problem of ensuring secure queries to the database PIR (Private Information Retrieval) problem. Previously, the authors considered the problem for a database hosted in the cloud in the presence of an active adversary who does not interfere with the execution of the protocol, but can carry out an attack with known open queries. In algorithms, bit number i is represented as the l-ary number with a number of digits d. An algorithm for placing a database in the cloud and an algorithm for querying the required bit using permutations in the digits of the bit number, using the specification of the bit number i in the base l numerical system, were proposed. Permutations are treated as secret encryption keys. Communication complexity and probability of guessing the bit number for a one-time attack with a known open request for bit number i and for an attack with unlimited number of known open requests were estimated.
12-23
DATA ANALYSIS
Adaptive IIR filter based on penalized spline
Abstract
The purpose of this research is to develop the technique of spline adaptive filters (SAF) for real-time implementation. The P-SAF proposed in the article based on the recurrent penalty P-spline, by analogy with the classical SAF, consists of linear dynamic and nonlinear static components. To adapt P-SAF, computing circuits with different topologies have been developed. This approach specifies a way to adapt the knots and calculate the spline coefficients simultaneously. This made it possible to increase the efficiency of P-SAF compared to the classical SAF and reduce computational costs. The efficiency indicator MSE [dB] for P-SAF is equal to and higher than for classical SAF when analyzing model and real time series.
24-34
Discrete optimization algorithm based on probability distribution with transformation of target values
Abstract
Optimization problems of searching in discrete space and, in particular, binary space, where a variable can take only two values, are of great practical importance. This paper proposes a new population discrete optimization algorithm based on probability distributions of variables. Distributions determine the probability of accepting one or another discrete value and are formed by transforming the target values of decisions into their weighting coefficients. The performance of the algorithm was assessed using unimodal and multimodal test functions with binary variables. The experimental results showed the high efficiency of the proposed algorithm in terms of convergence and stability estimates.
35-47
Optimization of software performance for classification and linking of administrative documents
Abstract
The paper discusses technologies for optimizing software performance. Optimization methods are divided into high-level and low-level, as well as parallelization. An algorithm for classifying and linking fields in a recognized image of an administrative document is described. The features of the implementation of classification and linking tasks are listed, consisting of the use of constellations of text feature points and the modified Levenshtein distance. SDK Smart Document Engine and OCR Tesseract were used. Several ways are described to optimize the performance of the functions for classifying and linking document content. Optimization of the performance of the system for sorting a stream of images of administrative documents is also described. The proposed methods for optimizing software performance are suitable not only for implementing image processing algorithms but also for computational algorithms in which cyclic information processing is carried out. The method can be applied in modern CAD systems to analyze the content of recognized textual files.
48-58