作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2023, Vol. 49 ›› Issue (3): 37-48. doi: 10.19678/j.issn.1000-3428.0064373

• 热点与综述 • 上一篇    下一篇

结合多解码器与两阶段通道选择的异常检测方法

王禹博, 陈利锋, 许卫霞   

  1. 复旦大学 计算机科学技术学院, 上海 200433
  • 收稿日期:2022-04-04 修回日期:2022-05-11 发布日期:2022-06-20
  • 作者简介:王禹博(1997—),男,硕士研究生,主研方向为深度异常检测;陈利锋,讲师、硕士;许卫霞,讲师、博士。
  • 基金资助:
    国家自然科学基金(U1636205);之江实验室开放基金(2019KB0AB05)。

Anomaly Detection Method Combining with Multi-Decoder and Two-Stage Channel Selection

WANG Yubo, CHEN Lifeng, XU Weixia   

  1. School of Computer Science, Fudan University, Shanghai 200433, China
  • Received:2022-04-04 Revised:2022-05-11 Published:2022-06-20

摘要: 异常检测是发现数据集中不符合明确定义的模式(或样本)的过程。目前流行的异常检测方法通常只对合群点进行建模,忽视了潜藏在无标记数据中的丰富信息,导致出现过拟合、阈值依赖等问题。此外,部分现有方法同等地处理正常类和异常类,与异常检测的目的相矛盾。提出一种结合多解码器与两阶段通道选择的异常检测方法,采用一种全新的重构-选择模型代替重构-排序-拒绝模型,并分别对合群点和离群点进行解码建模,构造一个无须阈值的选择器以区分两者,降低阈值确定过程中产生的人为误差和计算开销。通道选择器以两阶段的方式为样本分配通道,在第一阶段使用注意力选择器为无标记样本选择最佳匹配的离群点通道,在第二阶段使用竞争式选择器在合群点通道和最佳离群点通道中选择更适合的一个作为目标通道。两阶段选择器的结构能避免出现局部极值点,并可以利用直接引入的监督信息对选择结果进行修正。实验结果表明,与U-Std、ARAE、AnoGAN等流行方法相比,该方法的检测性能更好,在MNIST、Fashion-MNIST、CIFAR-10数据集上的AUROC值分别为99.3%、96.9%和90.5%。

关键词: 异常检测, 非对称建模, 数据增强, 数据集清理, 免阈值

Abstract: Anomaly detection is the process of finding patterns(or samples) in data that do not conform to well-defined behavior. Current popular anomaly detection methods always merely model the inliers, whilst ignoring the rich information that is contained in unlabeled data, leading to problems such as overfitting and threshold dependence. Meanwhile several methods treat normal and abnormal classes equally, which contradicts with the purpose of anomaly detection. A method for anomaly detection combining multi-decoder and two-stage channel selection is proposed. This technique adopts a brand new reconstruct-and-select framework to replace reconstruct-rank-reject framework, models Inliers and Outliers by decoding respectively, whilst building a threshold-free selector to distinguish amongst them, diminishing the human error and computational cost of threshold determination. The channel selector works in a two-stage fashion:at the first tage, the attentive selector chooses the best matching outlier channel for unlabeled sample;at the second stage, the competing selector decides whether the better target channel is Inlier channel or Outlier channel. The two-stage structure of selector makes it avoid local extremum and utilize directly introduced supervision information to correct the selection result. Experimental results demonstrate the detection performance of this method is better than U-Std, ARAE, AnoGAN and other popular methods, with the Area Under the Receiver Operating Characteristic Curve (AUROC) value on classes of MNIST, Fashion-MNIST and CIFAR-10 datasets reaching 99.3%, 96.9% and 90.5% respectively.

Key words: anomaly detection, asymmetric modeling, data augmentation, dataset cleaning, threshold-free

中图分类号: