ISSN 1000-1239 CN 11-1777/TP

Journal of Computer Research and Development ›› 2020, Vol. 57 ›› Issue (9): 2001-2008.doi: 10.7544/issn1000-1239.2020.20190462

Previous Articles    

Defending Against Dimensional Saddle Point Attack Based on Adaptive Method with Dynamic Bound

Li Dequan1, Xu Yue1, Xue Sheng2   

  1. 1(School of Mathematics and Big Data, Anhui University of Science and Technology, Huainan, Anhui 232001);2(School of Energy and Security, Anhui University of Science and Technology, Huainan, Anhui 232001)
  • Online:2020-09-01
  • Supported by: 
    This work was supported by the National Key Research and Development Program of China(2018YFF0301000), the National Natural Science Foundation of China (61472003), the Academic and Technical Leaders and the Backup Candidates of Anhui Province (2019H211), and the Program of the Innovation Team of “50 Star of Science and Technology” of Huainan, Anhui Province.

Abstract: With the advent of the era of big data, distributed machine learning has been widely applied to process massive data. The most commonly used one is the distributed stochastic gradient descent algorithm, but it is vulnerable to different types of Byzantine attacks. In order to maximize the elastic limit to defend against attacks and optimize objective function in the distributed dimensional Byzantine environment based on the gradient update rule, firstly a new Byzantine attack method—saddle point attack is proposed in this paper. Contrasting with the adaptive non-adaptive methods, the adaptation with dynamic bound escapes the saddle point fast when the objective function is stuck in the saddle point. The comparative experiment is made on the classification of data sets. Secondly, an aggregation rule Saddle(·) for filtering Byzantine agents is proposed, and it is proved that the rule is the dimensional Byzantine resilience. Therefore, in the distributed dimensional Byzantine environment, the adaptive optimization method with dynamic bound combined with the aggregation rule Saddle(·) can effectively defend against the saddle point attack. Finally, the error rate of the data set classification in the experimental results is compared to analyze the advantages and disadvantages of the adaptation with dynamic bound over the adaptive and non-adaptive methods. The result shows that the adaptation with dynamic bound combined with the aggregation rule Saddle(·) is less affected by the saddle point attack in the distributed dimensional Byzantine environment.

Key words: distributed optimization, dimensional Byzantine, saddle point attack, the adaption with dynamic bound, the aggregation rule Saddle(·)

CLC Number: