TY - JOUR

T1 - Multi-label active learning from crowds for secure IIoT

AU - Wu, Ming

AU - Li, Qianmu

AU - Bilal, Muhammad

AU - Xu, Xiaolong

AU - Zhang, Jing

AU - Hou, Jun

PY - 2021/10/1

Y1 - 2021/10/1

N2 - With the development of IIoT (Industrial Internet of Things), Artificial Intelligence technology is widely used in many research areas, such as image classification, speech recognition, and information retrieval. Traditional supervised machine learning obtains labels from high-quality oracles, which is high cost and time-consuming and does not consider security. Since multi-label active learning becomes a hot topic, it is more challenging to train efficient and secure classification models, and reduce the label cost in the field of IIoT. To address this issue, this research focuses on the secure multi-label active learning for IIoT using an economical and efficient strategy called crowdsourcing, which involves querying labels from multiple low-cost annotators with various expertise on crowdsourcing platforms rather than relying on a high-quality oracle. To eliminate the effects of annotation noise caused by imperfect annotators, we propose the Multi-label Active Learning from Crowds (MALC) method, which uses a probabilistic model to simultaneously compute the annotation consensus and estimate the classifier's parameters while also taking instance similarity into account. Then, to actively choose the most informative instances and labels, as well as the most reliable annotators, an instance-label-annotator triplets selection technique is proposed. Experimental results on two real-world data sets show that the performance of MALC is superior to existing methods.

AB - With the development of IIoT (Industrial Internet of Things), Artificial Intelligence technology is widely used in many research areas, such as image classification, speech recognition, and information retrieval. Traditional supervised machine learning obtains labels from high-quality oracles, which is high cost and time-consuming and does not consider security. Since multi-label active learning becomes a hot topic, it is more challenging to train efficient and secure classification models, and reduce the label cost in the field of IIoT. To address this issue, this research focuses on the secure multi-label active learning for IIoT using an economical and efficient strategy called crowdsourcing, which involves querying labels from multiple low-cost annotators with various expertise on crowdsourcing platforms rather than relying on a high-quality oracle. To eliminate the effects of annotation noise caused by imperfect annotators, we propose the Multi-label Active Learning from Crowds (MALC) method, which uses a probabilistic model to simultaneously compute the annotation consensus and estimate the classifier's parameters while also taking instance similarity into account. Then, to actively choose the most informative instances and labels, as well as the most reliable annotators, an instance-label-annotator triplets selection technique is proposed. Experimental results on two real-world data sets show that the performance of MALC is superior to existing methods.

KW - Active learning

KW - Annotation consensus

KW - Crowdsourcing

KW - Multi-label learning

KW - Secure IIoT

U2 - 10.1016/j.adhoc.2021.102594

DO - 10.1016/j.adhoc.2021.102594

M3 - Journal article

AN - SCOPUS:85108909129

VL - 121

JO - Ad Hoc Networks

JF - Ad Hoc Networks

SN - 1570-8705

M1 - 102594

ER -