Dynamic constraint adaptive method based on aggregation rule

Abstract

In a distributed high-dimensional Byzantine environment, a dynamically constrained adaptive optimization method is provided and is combined with an algorithm of an aggregation rule Saddle (.), so thatthe deliberate attack problem can be resisted to the greatest extent elastically, and the optimization problem can be effectively solved. On the basis of a gradient updating rule, firstly, a new Byzantine attack mode, namely saddle point attack, is provided. And when the target function is caught in the saddle point, compared with self-adaptive and non-self-adaptive methods, the provided dynamicconstraint self-adaptive method can escape from the saddle point more quickly. Secondly, an aggregation rule Saddle (.) for filtering a Byzantine individual is constructed, and in a distributed high-dimensional Byzantine environment, a dynamic constraint adaptive optimization method is combined with the aggregation rule Saddle (.), so that saddle point attacks can be effectively resisted. Finally,the advantages and disadvantages of dynamic constraint adaptive and adaptive and non-adaptive methods are verified, compared and analyzed through experiments. Results show that the dynamic constraintself-adaption combined with the aggregation rule Saddle (.) is less affected by saddle point attacks in a distributed high-dimensional Byzantine environment.

Description

technical field [0001] The invention relates to an adaptive method for dynamic constraints based on aggregation rules, which belongs to the field of machine learning. Background technique [0002] With the rapid development of sensor and smart device technology and its wide application in practice, it has become a reality to collaboratively collect a large amount of data from multiple information sources. Faced with the processing of massive data with spatial distribution, as the main effective tool capable of extracting valuable information from it, machine learning has achieved great success in applications in a wide range of fields such as computer vision, healthcare, and financial market analysis. Among them, the distributed stochastic gradient descent algorithm is one of the most commonly used machine learning methods. It only randomly trains one sample data in each iteration, which greatly reduces the training time. But it is vulnerable to Byzantine attacks. Errors i...

AI Technical Summary

Problems solved by technology

[0005] The technical problem to be solved by the present invention is to provide a dynamic constraint adaptive method based on aggregation rules, aiming to solve the optimization problem of maximum flexibility to resist attacks in a distributed high-dimensional Byzantine environment

Images