异常检测，GAN如何gan ?

     欢迎点击上方蓝字，关注啦 ~

相关阅读：

【1】GAN在医学图像上的生成，今如何？

虚拟换衣！速览这几篇最新论文咋做的！

脸部妆容迁移！速览几篇用GAN来做的论文

GAN整整6年了！是时候要来捋捋了！

01-GAN公式简明原理之铁甲小宝篇

今天记录一下、一些用GAN来做异常检测的论文！

---

异常检测(Anomaly detection)，一个很常见的问题。

在图像方面，比如每天出入地铁安检，常常看到小姐姐小哥哥们坐在那盯着你的行李过检图像，类似如下（图来自GANomaly论文）：

又比如在一些医学图像分析上，源自健康人的影像也许是比较容易获取的，并且图像的“模式”往往固定或者不多变的，而病变的图像数量是很少、很难获取，或者病变区域多变、甚至未知的，此时异常检测就面临着正样本/异常图像很少，而相对地，正常图像更容易获得的情况。这种情况其实在很多场景下有所体现，比如工业视觉检测等等。

对于已知类别、数量较多情况下，不管异常与否，我们也许可以通过训练一个分类模型就能解决。但面对也许未知、多变的情况，要想用一个多分类模型分辨出来似乎很难。如果是想仅仅分辨出是不是异常，那也许可以做一个单分类器即可。

我们尽可能地去让模型充分学习正常数据的分布长什么样子，一旦来了异常图像，它即便不知道这是啥新的分布，但依旧可以自信地告诉你：这玩意儿没见过，此乃异类也！

用GAN一些网络怎么做呢？大体思想是：

在仅有负样本（正常数据）或者少量正样本情况下：

训练阶段：

      可以通过网络仅仅学习负样本（正常数据）的数据分布，得到的模型G只能生成或者重建正常数据。

测试阶段：

      使用测试样本输入训练好的模型G，如果G经过重建后输出和输入一样或者接近，表明测试的是正常数据，否则是异常数据。

模型G的选择：

      一个重建能力或者学习数据分布能力较好的生成模型，例如GAN或者VAE，甚至encoder-decoder。

下面速览几篇论文、看看GAN是如何做异常检测的（数据主要为图像形式）：

---

1. IPMI 2017 AnoGAN ( Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery )

思路：通过一个GAN的生成器G来学习正常数据的分布，测试时图像通过学习到的G找到它应该的正常图的样子，再通过对比来找到异常与否的情况。

如上图所示，AnoGAN论文中采用的是DCGAN，一种较简单的GAN架构。

训练阶段：

      对抗训练，从一个噪声向量Z通过几层反卷积搭建的生成器G学习生成正常数据图像。

测试阶段：

      随机采样一个高斯噪声向量z，想要通过已经训练好的G生成一幅和测试图像x对应的正常图像G(z)。G的参数是固定的，它只能生成落在正常数据分布的图像。但此时仍需进行训练，把z看成待更新的参数，通过比较G(z)和x的差异去更新，从而生成一个与x尽可能相似、理想对应的正常图像。

如果x是正常的图像，那么x和G(z)应该是一样的。

如果x异常，通过更新z，可以认为重建出了异常区域的理想的正常情况，这样两图一对比不仅仅可以认定异常情况，同时还可以找到异常区域。

为了比较G(z)和x差异去更新z：

一是通过计算G(z)和x的图像层面的L1 loss：

二是利用到训练好的判别器D，取G(z)和x在判别器D的中间层的特征层面的loss：

两者综合：

另外，异常分数计算方法：

2. 2018-02 EFFICIENT GAN-BASED ANOMALY DETECTION

针对AnoGAN测试阶段仍然需要更新参数的缺陷，此方法提出一种基于BiGAN可快百倍的方法。

训练时，同时学习将输入样本x映射到潜在表示z的编码器E，以及生成器G和判别器D:

如此可避免测试仍需要“找到z”那个耗时的步骤。与常规GAN中的D仅考虑输入（实际的或生成的）图像不同，而还考虑了潜在表示z(作为输入)。

测试时，判断图像的异常与否的分值计算方法，可选择可AnoGAN基本一样的方法。

3. 2018-12 Adversarially Learned Anomaly Detection

第二种方法的加强版，也是基于BiGAN,并且在稳定训练上做了些功夫。如下所示，（乖乖，搞了三个判别器 =_=

检测时的计算方法：

4. 2018-11-13 GANomaly: Semi-Supervised Anomaly Detection via Adversarial Training

原理：

训练时，约束正常的数据编码得到潜在空间表示z1，和对z1解码、再编码得到的z2，差距不会特别大，理想应该是一样的。

所以训练好后，用正常样本训练好的 G只能重建正常数据分布，一旦用于从未见过的异常样本编码、解码、再经历编码得到的潜在空间Z差距是大的。

当两次编码得到的潜在空间差距大于一定阈值的时候，就判定样本是异常样本。

5. 2019-01-25 Skip-GANomaly: Skip Connected and Adversarially Trained Encoder-Decoder Anomaly Detection

6. PRICAI 2018 A Surface Defect Detection Method Based on Positive Samples

原理：

C（x~|x）是人工缺陷制造模块。X~是模拟缺陷的样本，经过EN-DE编码解码器后重建正常样本Y。

测试阶段，X输入EN-DE后得到理想正常样本y，使用LBP对Y和X逐像素特征比较，相差大则有缺陷。

7. MIDL 2018  Unsupervised Detection of Lesions in Brain MRI using constrained adversarial auto-encoders

使用的是AAE来学习建模正常数据分布。有时，对于在正常分布的的两个数据之间的距离，比一个正常和一个异常之间的距离还大，所以提出在隐空间也加一个约束。

暂时先写到这吧。

最后，欢迎关注公众号啦~

最后的最后，再来发一波、到目前为止、部分、用 GAN做异常检测的基本相关（直接用“adversarial anomaly detection”在arxiv上爬下来的，不一定相关！2333）论文供参考！！！

---

 

001 (2019-12-10) Event Detection in Micro-PMU Data  A Generative Adversarial Network ScoringMethod

    https://arxiv.xilesou.top/pdf/1912.05103.pdf

 

002 (2019-12-10) Outage Detection in Partially Observable DistributionSystems using Smart Meters and Generative Adversarial Networks

    https://arxiv.xilesou.top/pdf/1912.04992.pdf

 

003 (2019-12-9) Oversampling Log Messages Using a Sequence GenerativeAdversarial Network for Anomaly Detection and Classification

    https://arxiv.xilesou.top/pdf/1912.04747.pdf

 

004 (2019-12-2) Anomaly Detection in Particulate Matter Sensor usingHypothesis Pruning Generative Adversarial Network

    https://arxiv.xilesou.top/pdf/1912.00583.pdf

 

005 (2019-11-27) Sparse-GAN Sparsity-constrained Generative Adversarial Network for AnomalyDetection in Retinal OCT Image

    https://arxiv.xilesou.top/pdf/1911.12527.pdf

 

006 (2019-11-21) EvAn  NeuromorphicEvent-based Anomaly Detection

    https://arxiv.xilesou.top/pdf/1911.09722.pdf

 

007 (2019-11-19) Attention Guided Anomaly Detection and Localization in Images

    https://arxiv.xilesou.top/pdf/1911.08616.pdf

 

008 (2019-11-17) Deep Verifier Networks Verification of Deep Discriminative Models with Deep Generative Models

    https://arxiv.xilesou.top/pdf/1911.07421.pdf

 

009 (2019-11-16) RSM-GAN  A ConvolutionalRecurrent GAN for Anomaly Detection in Contaminated Seasonal Multivariate TimeSeries

    https://arxiv.xilesou.top/pdf/1911.07104.pdf

 

010 (2019-10-30) Robust and Computationally-Efficient Anomaly Detectionusing Powers-of-Two Networks

    https://arxiv.xilesou.top/pdf/1910.14096.pdf

 

011 (2019-10-29) Small-GAN  SpeedingUp GAN Training Using Core-sets

    https://arxiv.xilesou.top/pdf/1910.13540.pdf

 

012 (2019-10-23) Photoshopping Colonoscopy Video Frames

    https://arxiv.xilesou.top/pdf/1910.10345.pdf

 

013 (2019-10-21) GraphSAC  Detectinganomalies in large-scale graphs

    https://arxiv.xilesou.top/pdf/1910.09589.pdf

 

014 (2019-10-21) Adversarial Anomaly Detection for Marked Spatio-TemporalStreaming Data

    https://arxiv.xilesou.top/pdf/1910.09161.pdf

 

015 (2019-10-10) Misbehaviour Prediction for Autonomous Driving Systems

    https://arxiv.xilesou.top/pdf/1910.04443.pdf

 

016 (2019-10-9) Adversarial Learning of Deepfakes in Accounting

    https://arxiv.xilesou.top/pdf/1910.03810.pdf

 

017 (2019-09-12) Perceptual Image Anomaly Detection

    https://arxiv.xilesou.top/pdf/1909.05904.pdf

 

018 (2019-08-27) Self-Supervised Representation Learning viaNeighborhood-Relational Encoding

    https://arxiv.xilesou.top/pdf/1908.10455.pdf

 

019 (2019-08-10) Transcriptional Response of SK-N-AS Cells to Methamidophos

    https://arxiv.xilesou.top/pdf/1908.03841.pdf

 

020 (2019-09-3) Februus  InputPurification Defence Against Trojan Attacks on Deep Neural Network Systems

    https://arxiv.xilesou.top/pdf/1908.03369.pdf

 

021 (2019-08-8) What goes around comes around  Cycle-Consistency-based Short-Term MotionPrediction for Anomaly Detection using Generative Adversarial Networks

    https://arxiv.xilesou.top/pdf/1908.03055.pdf

 

022 (2019-08-2) Detection of Accounting Anomalies in the Latent Space usingAdversarial Autoencoder Neural Networks

    https://arxiv.xilesou.top/pdf/1908.00734.pdf

 

023 (2019-09-3) Q-MIND  DefeatingStealthy DoS Attacks in SDN with a Machine-learning based Defense Framework

    https://arxiv.xilesou.top/pdf/1907.11887.pdf

 

024 (2019-10-8) Real-time Evasion Attacks with Physical Constraints on DeepLearning-based Anomaly Detectors in Industrial Control Systems

    https://arxiv.xilesou.top/pdf/1907.07487.pdf

 

025 (2019-07-12) AMAD  AdversarialMultiscale Anomaly Detection on High-Dimensional and Time-Evolving CategoricalData

    https://arxiv.xilesou.top/pdf/1907.06582.pdf

 

026 (2019-06-27) A Survey on GANs for Anomaly Detection

    https://arxiv.xilesou.top/pdf/1906.11632.pdf

 

027 (2019-06-15) Physical Integrity Attack Detection of Surveillance Camerawith Deep Learning Based Video Frame Interpolation

    https://arxiv.xilesou.top/pdf/1906.06475.pdf

 

028 (2019-07-8) GAN-based Multiple Adjacent Brain MRI Slice Reconstructionfor Unsupervised Alzheimer's Disease Diagnosis

    https://arxiv.xilesou.top/pdf/1906.06114.pdf

 

029 (2019-06-3) Generative Adversarial Networks for Distributed IntrusionDetection in the Internet of Things

    https://arxiv.xilesou.top/pdf/1906.00567.pdf

 

030 (2019-11-20) Unsupervised Learning of Anomaly Detection fromContaminated Image Data using Simultaneous Encoder Training

    https://arxiv.xilesou.top/pdf/1905.11034.pdf

 

031 (2019-10-18) Adversarially-trained autoencoders for robust unsupervisednew physics searches

    https://arxiv.xilesou.top/pdf/1905.10384.pdf

 

032 (2019-05-19) Spatio-Temporal Adversarial Learning for Detecting UnseenFalls

    https://arxiv.xilesou.top/pdf/1905.07817.pdf

 

033 (2019-05-20) Finding Rats in Cats Detecting Stealthy Attacks using Group Anomaly Detection

    https://arxiv.xilesou.top/pdf/1905.07273.pdf

 

034 (2019-04-25) End-to-End Adversarial Learning for Intrusion Detection inComputer Networks

    https://arxiv.xilesou.top/pdf/1904.11577.pdf

 

035 (2019-04-24) GAN Augmented Text Anomaly Detection with Sequences of DeepStatistics

    https://arxiv.xilesou.top/pdf/1904.11094.pdf

 

036 (2019-04-23) A Comparison Study of Credit Card Fraud Detection  Supervised versus Unsupervised

    https://arxiv.xilesou.top/pdf/1904.10604.pdf

 

037 (2019-09-24) Trick or Heat Manipulating Critical Temperature-Based Control Systems UsingRectification Attacks

    https://arxiv.xilesou.top/pdf/1904.07110.pdf

 

038 (2019-12-2) Adversarial Learning in Statistical Classification  A Comprehensive Review of Defenses AgainstAttacks

    https://arxiv.xilesou.top/pdf/1904.06292.pdf

 

039 (2019-04-11) (Martingale) Optimal Transport And Anomaly Detection WithNeural Networks  A Primal-dual Algorithm

    https://arxiv.xilesou.top/pdf/1904.04546.pdf

 

040 (2019-07-24) Efficient GAN-based method for cyber-intrusion detection

    https://arxiv.xilesou.top/pdf/1904.02426.pdf

 

041 (2019-04-2) Fence GAN  TowardsBetter Anomaly Detection

    https://arxiv.xilesou.top/pdf/1904.01209.pdf

 

042 (2019-03-27) Fundamental Limits of Covert Packet Insertion

    https://arxiv.xilesou.top/pdf/1903.11640.pdf

 

043 (2019-05-20) Deep Generative Design Integration of Topology Optimization and Generative Models

    https://arxiv.xilesou.top/pdf/1903.01548.pdf

 

044 (2019-11-14) adVAE  Aself-adversarial variational autoencoder with Gaussian anomaly prior knowledgefor anomaly detection

    https://arxiv.xilesou.top/pdf/1903.00904.pdf

 

045 (2019-07-14) Secure Distributed Dynamic State Estimation in Wide-AreaSmart Grids

    https://arxiv.xilesou.top/pdf/1902.07288.pdf

 

046 (2019-02-19) Anomaly Detection with Adversarial Dual Autoencoders

    https://arxiv.xilesou.top/pdf/1902.06924.pdf

 

047 (2019-05-9) The Odds are Odd  AStatistical Test for Detecting Adversarial Examples

    https://arxiv.xilesou.top/pdf/1902.04818.pdf

 

048 (2019-11-6) BIVA  A Very DeepHierarchy of Latent Variables for Generative Modeling

    https://arxiv.xilesou.top/pdf/1902.02102.pdf

 

049 (2019-01-28) Heartbeat Anomaly Detection using Adversarial Oversampling

    https://arxiv.xilesou.top/pdf/1901.09972.pdf

 

050 (2019-01-25) Skip-GANomaly  SkipConnected and Adversarially Trained Encoder-Decoder Anomaly Detection

    https://arxiv.xilesou.top/pdf/1901.08954.pdf

 

051 (2019-05-27) Maximum Entropy Generators for Energy-Based Models

    https://arxiv.xilesou.top/pdf/1901.08508.pdf

 

052 (2019-01-10) Adversarial Pseudo Healthy Synthesis Needs PathologyFactorization

    https://arxiv.xilesou.top/pdf/1901.07295.pdf

 

053 (2019-01-18) Robust Anomaly Detection in Images using AdversarialAutoencoders

    https://arxiv.xilesou.top/pdf/1901.06355.pdf

 

054 (2019-01-15) MAD-GAN  MultivariateAnomaly Detection for Time Series Data with Generative Adversarial Networks

    https://arxiv.xilesou.top/pdf/1901.04997.pdf

 

055 (2019-12-4) Event Generation and Statistical Sampling for Physics withDeep Generative Models and a Density Information Buffer

    https://arxiv.xilesou.top/pdf/1901.00875.pdf

 

056 (2018-12-11) Anomaly Generation using Generative Adversarial Networks inHost Based Intrusion Detection

    https://arxiv.xilesou.top/pdf/1812.04697.pdf

 

057 (2018-12-11) Anomaly detection with Wasserstein GAN

    https://arxiv.xilesou.top/pdf/1812.02463.pdf

 

058 (2018-12-5) Adversarially Learned Anomaly Detection

    https://arxiv.xilesou.top/pdf/1812.02288.pdf

 

059 (2018-11-11) Adversarial Learning-Based On-Line Anomaly Monitoring forAssured Autonomy

    https://arxiv.xilesou.top/pdf/1811.04539.pdf

 

060 (2018-10-19) Subset Scanning Over Neural Network Activations

    https://arxiv.xilesou.top/pdf/1810.08676.pdf

 

061 (2018-10-11) MDGAN  BoostingAnomaly Detection Using \\Multi-Discriminator Generative Adversarial Networks

    https://arxiv.xilesou.top/pdf/1810.05221.pdf

 

062 (2019-04-30) Prospect Theoretic Approach for Data Integrity in IoTNetworks under Manipulation Attacks

    https://arxiv.xilesou.top/pdf/1809.07928.pdf

 

063 (2019-01-15) Anomaly Detection with Generative Adversarial Networks forMultivariate Time Series

    https://arxiv.xilesou.top/pdf/1809.04758.pdf

 

064 (2018-09-28) Layerwise Perturbation-Based Adversarial Training for HardDrive Health Degree Prediction

    https://arxiv.xilesou.top/pdf/1809.04188.pdf

 

065 (2018-09-7) Coupled IGMM-GANs for deep multimodal anomaly detection inhuman mobility data

    https://arxiv.xilesou.top/pdf/1809.02728.pdf

 

066 (2019-08-2) Detection and Mitigation of Attacks on TransportationNetworks as a Multi-Stage Security Game

    https://arxiv.xilesou.top/pdf/1808.08349.pdf

 

067 (2018-08-23) DOPING  GenerativeData Augmentation for Unsupervised Anomaly Detection with GAN

    https://arxiv.xilesou.top/pdf/1808.07632.pdf

 

068 (2018-08-1) Anomaly Detection via Minimum Likelihood GenerativeAdversarial Networks

    https://arxiv.xilesou.top/pdf/1808.00200.pdf

 

069 (2018-07-22) SAIFE  UnsupervisedWireless Spectrum Anomaly Detection with Interpretable Features

    https://arxiv.xilesou.top/pdf/1807.08316.pdf

 

070 (2018-06-27) Adversarial Distillation of Bayesian Neural NetworkPosteriors

    https://arxiv.xilesou.top/pdf/1806.10317.pdf

 

071 (2019-03-25) Learning Neural Random Fields with Inclusive AuxiliaryGenerators

    https://arxiv.xilesou.top/pdf/1806.00271.pdf

 

072 (2018-07-17) AVID  AdversarialVisual Irregularity Detection

    https://arxiv.xilesou.top/pdf/1805.09521.pdf

 

073 (2018-11-13) GANomaly Semi-Supervised Anomaly Detection via Adversarial Training

    https://arxiv.xilesou.top/pdf/1805.06725.pdf

 

074 (2018-05-5) Population Anomaly Detection through Deep Gaussianization

    https://arxiv.xilesou.top/pdf/1805.02123.pdf

 

075 (2018-04-13) Group Anomaly Detection using Deep Generative Models

    https://arxiv.xilesou.top/pdf/1804.04876.pdf

 

076 (2018-04-13) Adversarial Clustering A Grid Based Clustering Algorithm Against Active Adversaries

    https://arxiv.xilesou.top/pdf/1804.04780.pdf

 

077 (2018-04-12) Deep Autoencoding Models for Unsupervised AnomalySegmentation in Brain MR Images

    https://arxiv.xilesou.top/pdf/1804.04488.pdf

 

078 (2018-04-3) Correlated discrete data generation using adversarialtraining

    https://arxiv.xilesou.top/pdf/1804.00925.pdf

 

079 (2018-03-17) A Multi-perspective Approach To Anomaly Detection ForSelf-aware Embodied Agents

    https://arxiv.xilesou.top/pdf/1803.06579.pdf

 

080 (2018-04-9) CIoTA  CollaborativeIoT Anomaly Detection via Blockchain

    https://arxiv.xilesou.top/pdf/1803.03807.pdf

 

081 (2018-05-24) Adversarially Learned One-Class Classifier for NoveltyDetection

    https://arxiv.xilesou.top/pdf/1802.09088.pdf

 

082 (2019-05-1) Efficient GAN-Based Anomaly Detection

    https://arxiv.xilesou.top/pdf/1802.06222.pdf

 

083 (2018-02-13) Satellite Image Forgery Detection and Localization UsingGAN and One-Class Classifier

    https://arxiv.xilesou.top/pdf/1802.04881.pdf

 

084 (2018-02-8) Detection of Adversarial Training Examples in PoisoningAttacks through Anomaly Detection

    https://arxiv.xilesou.top/pdf/1802.03041.pdf

 

085 (2018-01-5) Shielding Google's language toxicity model againstadversarial attacks

    https://arxiv.xilesou.top/pdf/1801.01828.pdf

 

086 (2018-06-27) When Not to Classify Anomaly Detection of Attacks (ADA) on DNN Classifiers at Test Time

    https://arxiv.xilesou.top/pdf/1712.06646.pdf

 

087 (2018-04-24) Bayesian Hypernetworks

    https://arxiv.xilesou.top/pdf/1710.04759.pdf

 

088 (2017-09-15) To Go or Not To Go  ANear Unsupervised Learning Approach For Robot Navigation

    https://arxiv.xilesou.top/pdf/1709.05439.pdf

 

089 (2017-04-5) Counter-RAPTOR Safeguarding Tor Against Active Routing Attacks

    https://arxiv.xilesou.top/pdf/1704.00843.pdf

 

090 (2017-03-17) Unsupervised Anomaly Detection with Generative AdversarialNetworks to Guide Marker Discovery

    https://arxiv.xilesou.top/pdf/1703.05921.pdf