基于少样本深度学习的异常检测方法:综述与展望

Survey on Few-shot Deep Learning Based Anomaly Detection

随着深度学习技术的迅猛发展和广泛应用,需要进行异常检测的场景也越来越多.然而,在对深度神经网络进行训练时,现有的数据集往往不足够支撑模型进行有效的训练,应用少样本学习进行异常检测获得了广泛的关注.本文首先对异常检测问题进行了说明,对少样本学习进行异常检测进行了定义;其次,本文系统的总结了基于少样本学习进行异常检测的方法,对基于少样本学习进行异常检测从两个角度进行了分类,总结了评价指标,介绍了常用的数据集;最后,本文对少样本学习异常检测的未来进行了展望,提出了研究方向和思路.

With the rapid development and widespread application of deep learning technology,there is an increasing need for anomaly detection in various scenarios.However,during the training of deep neural networks,existing datasets often do not sufficiently support effective model training.As a result,the application of few-shot learning for anomaly detection has gained widespread attention.This paper first provides an explanation of the anomaly detection problem and defines anomaly detection using few-shot learning.Secondly,the paper systematically summarizes the methods of anomaly detection based on few-shot learning,categorizing anomaly detection based on few-shot learning from two perspectives,summarizing evaluation metrics,and introducing commonly used datasets.Finally,the paper provides an outlook on the future of few-shot learning based anomaly detection,proposing research directions and ideas.