Data Poisoning Attacks And Defense Against Truth Discovery In Crowdsensing Systems

Crowdsensing systems collect multidimensional and heterogeneous data by sensing devices of individuals,solving the problem of large-scale data needs.However,there are often conflicts between data from multiple sources,many crowdsensing services also lack of strict authentication for user identity,which means attackers can sneak into normal workers,and submit malicious data to launch data poisoning attacks.How to find the truth from conflicting information has become an important issue.Truth discovery aim to calculate the trustworthiness of data sources and try to find ground truth from inconsistent data.It can help the system filter out poor quality data providers,and defend against simple poisoning strategies,such as random attacks,max-value attacks,etc.However,in multi-rounds of data collection tasks,attackers may launch an effective attack by slightly modifying simple poisoning strategies.Attackers deceive truth discovery by alternately submitting real and fake data,thereby misleading the system to infer the wrong truth.This attack strategy may lead to longer-term,more insidious,and more damaging poisoning attacks,affecting the proper operation of crowdsensing system.It is necessary to carry out data poisoning attacks and defense work against truth discovery in crowdsensing scenarios.The main work includes the following parts:(1)This work focuses on numerical sensing tasks and partially observable multi-round data poisoning attack scenarios.This work demonstrates the necessity of truth discovery work for conflicting data,consider crowdsensing system as a gray box,and use deep reinforcement learning to model multi-round data poisoning attack scenarios,which can batch train malicious workers to make decisions automatically.(2)This work make optimizations from two defense perspectives,malicious worker detection and high robust truth discovery,propose a High Robust Truth Discovery with Malicious worker Detection(HRTDMD)that enables it to resist multi-round attacks.(3)This work designs an attack and defense system for truth discovery under crowdsensing system.The system supports users to perform attack and defense exercises in the test platform and visualize the attack and defense effects.