| With the development and popularization of smart grid,more and more terminal devices are connected to the grid network,but it becomes a difficult problem to solve the computing demand of terminal devices for real-time.The introduction of edge computing is a good solution,and the limitation of bandwidth,delay and other aspects of the power grid system previously relying only on cloud computing will be alleviated.However,edge computing is vulnerable to security aspects due to the limitations caused by computing power and storage resources.Among the multiple attack scenarios in edge computing,the attacks involving machine learning aspects are relatively less studied.As more and more important decisions in edge computing rely on machine learning models,it is necessary to worry about the potential threats posed by machine learning algorithms in edge computing environments.Poisoning attack is an important direction about machine learning security countermeasures research,the attacker adds carefully constructed poison sample data to the training dataset before training the machine learning model,at which time the target model obtained by the learning algorithm will be affected.For the characteristics of edge computing,machine learning at the edge side is usually set up as incremental learning and online learning,and incremental learning is the main target of poisoning attacks,so the research of poisoning attacks on edge computing is of great significance.In this thesis,a variety of poisoning attack strategies are proposed for a variety of online prediction scenarios in edge computing environment.The main contributions of this thesis are as follows:1.For the case of limited edge computing resources and the demand of power system at the edge side,this thesis uses online learning to build regression task models for the edge side and implements a variety of power prediction algorithms online in the edge computing environment.2.Attacks research on edge computing of smart grid is mainly focused on false data injection attacks,there is little research on poisoning attacks on edge-side machine learning models,and this thesis provides an idea and direction for related research.3.In the field of poisoning attack research,most of them are for classification tasks,there are only a few studies on poisoning attacks for regression tasks.The two types of poisoning attacks proposed in this thesis,white-box attacks and black-box attacks,are both research on poisoning attacks for regression tasks.4.This thesis is the study of poisoning attacks on regression tasks based on online learning.For the study of poisoning attacks on regression models,they are currently aimed at offline training methods,and the maximum loss attack and subsequence finding attack introduced in this thesis are both for the study of poisoning attacks on online regression tasks.5.This thesis proposes two black-box poisoning attacks for regression tasks,which are more widely used in practical scenarios and can produce attack effects on some complex target models.The several poisoning attack strategies based on edge computing environments proposed in this thesis demonstrate the feasibility of the attacks in experimental simulations and deploy the first line of defense for machine learning security decisions in edge computing environments. |
