Research on Local Differential Privacy Protection of High Dimensional Data in Embedded System Based on Hybrid Differential Swarm Algorithm

Shiru Sun ； Bin Wang ； Hongwei Zhang

《International Journal of Network Security》 25卷3期 (2023/05) Pp. 537-544

https://doi.org/10.6633/IJNS.202305_25(3).19

摘要

Deep neural networks have significant gradient redundancy in gradient descent. Therefore, excessive noise is introduced when the differential privacy mechanism is used to resist member inference attacks. This paper proposes a local differential privacy protection theme of high dimensional data in embedded systems based on a hybrid differential swarm algorithm to solve this problem. Firstly, the data source differential privacy protection algorithm is used to disturb the client data set and generate the disturbed data set to protect the privacy of the original local data set. Then, a hybrid differential swarm algorithm reduces the high-dimensional data set to multiple low-dimensional attribute sets, and a new data set is synthesized. The algorithm uses the maximum spanning tree criterion to get the initial population. Then the crossover and mutation rules in differential evolution algorithms optimize the initial population. Using the differential evolution algorithm, the swarm algorithm is applied to the mutation stage, the optimization and improvement crossover stage, respectively, and the adaptive cloud theory is applied to the selection stage to select the generated individuals. Finally, the proposed algorithm is verified on the standard data sets MNIST and CIFAR-10 to more effectively bridge the gap between the proposed algorithm and the non-privacy model.

關鍵字

Embedded System ； High Dimensional Data ； Hybrid Differential Swarm Algorithm ； Local Differential Privacy Protection