Anomalous Example Detection in Deep Learning: A Survey

Abstract

Deep Learning (DL) is vulnerable to out-of-distribution and adversarial examples resulting in incorrect outputs. To make DL more robust, several posthoc anomaly detection techniques to detect (and discard) these anomalous samples have been proposed in the recent past. This survey tries to provide a structured and comprehensive overview of the research on anomaly detection for DL based applications. We provide a taxonomy for existing techniques based on their underlying assumptions and adopted approaches. We discuss various techniques in each of the categories and provide the relative strengths and weaknesses of the approaches. Our goal in this survey is to provide an easier yet better understanding of the techniques belonging to different categories in which research has been done on this topic. Finally, we highlight the unsolved research challenges while applying anomaly detection techniques in DL systems and present some high-impact future research directions.

Adversarial Example Detection for DNN Models: A Review and Experimental Comparison

“…In [28], the authors were the first who classified the detection methods into: 1) auxiliary models in which a subnetwork or a separate network acts as classifier to predict adversarial inputs, 2) statistical models in which statistical analyses were used to distinguish between normal and adversarial inputs and 3) prediction consistency based models that depends on the model prediction if the input or the model parameters are changed. In the review of Bulusu et al [58] and of Miller et al [59,60], they classified detection methods with respect to the presence of AEs in the training process of the detector into: 1) supervised detection in which AEs are used in the training of the detector and 2) unsupervised detection in which the detector is only trained using normal training data. Carlini et al [25] did an experimental study on ten detectors to show that all can be defeated by building new loss functions.…”

A Comprehensive Survey on Graph Anomaly Detection with Deep Learning

“…Specifically, [42], [43] and [41] are representative surveys on the generalized anomaly detection techniques. But only the most up-to-date work in Thudumu et al [41] covers the topic of graph anomaly detection.…”

A Critical Study on the Recent Deep Learning Based Semi-Supervised Video Anomaly Detection Methods

“…Moreover, this review explains the basic assumption, advantages, computational cost, etc, for each of the techniques. Bulusu [11] has provided a review on DL-based anomalous instance detection methods. Its focus is on discussing unintentional and intentional anomalies, specifically in the context of DNNs.…”