Anomaly detection using Autoencoders and Deep Convolution Generative Adversarial Networks

摘要

The current trend in transport (road, railway, unmanned aerial vehicle) introduces autonomous transport means. As the autonomy of vehicles increases, situations arise for which the models have not been trained, which increases the safety risk. Therefore, we propose systems that would detect such “anomaly” situations. We designed two anomaly detectors - an Adversarial Autoencoder (AAE) and a Deep Convolutional Generative Adversarial Networks (DCGAN). These models are build up on models from resources Autoencoders (2020) and Deep (2020). Networks are trained using picture datasets MNIST, Fashion-MNIST and CIFAR10.With DCGAN network, cumulative and reverse cumulative distribution functions are used to find an optimal decision threshold. In the case of autoencoder networks – an optimal number of latent variables is found using reconstruction error ratio function. Then, in both cases (AAE and DCGAN network), cumulative and reverse cumulative distribution functions are used to find an optimal decision threshold. Finally, influence of image picture complexity on the anomaly detection is discussed. We got best results with anomaly detectors trained on the less complex datasets comparing to test datasets. For both detectors trained on the simplest database MNIST, under a given anomaly expectation probability equal to 0.5, we reached the anomaly detection error 0.08% (AAE) and 1.89% (DCGAN).