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）：

 

异常检测，GAN如何gan ?_其他

 

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！

异常检测，GAN如何gan ?_其他_02

 

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。

 

异常检测，GAN如何gan ?_其他_03

 

 

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：

异常检测，GAN如何gan ?_其他_04

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：

异常检测，GAN如何gan ?_其他_05

两者综合：

异常检测，GAN如何gan ?_其他_06

另外，异常分数计算方法：

异常检测，GAN如何gan ?_其他_07

 

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:

异常检测，GAN如何gan ?_其他_08

异常检测，GAN如何gan ?_其他_09

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。

异常检测，GAN如何gan ?_其他_10

 

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 =_=

异常检测，GAN如何gan ?_其他_11

检测时的计算方法：

异常检测，GAN如何gan ?_其他_12

 

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

异常检测，GAN如何gan ?_其他_13

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。

异常检测，GAN如何gan ?_其他_14

 

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

异常检测，GAN如何gan ?_其他_15

 

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

异常检测，GAN如何gan ?_其他_16

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。

 

异常检测，GAN如何gan ?_其他_17

 

暂时先写到这吧。

 

最后，欢迎关注公众号啦~

异常检测，GAN如何gan ?_其他_18

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