Anomaly detection with ML and deep neural networks

Based on the existing data of fraudsters and ordinary users, we have utilized deep neural networks technology to define certain patterns of activities performed by fraud users. The determined chain of actions got gathered into a pool of fraud behavior. Using machine learning, we are able to prepare a certain model for learning, input the gathered fraud pattern, and further analyze the predicted user behavior. It is impossible to so with standard analytical tools.

The central part of this whole process is neural network model development. We create various algorithms for neural networks based on TensorFlow, SK Learn, Keras, and other narrowly focused ML frameworks. A combination of those frameworks allowed us to process huge amounts of data available in this social media, to learn under high loads, and analyze information. In general, we've used a variety of neural network architectures like recurrent and convolutional neural networks. Then we've launched each developed model, tested and fine-tuned its parameters to make it ready for further data predictions. The process of these models’ development is continuous and relies on up-to-date data.

The developed by Uinno ensemble of neural network models allows allocating fraud users based on the defined pool of fraud actions within a few days after registration. The system gathers certain characteristics from the behavior of new users, defines the most appropriate analysis algorithm from the existing ones, stores it into a specific bucket list with Amazon S3. Then the information is processed using the developed neural network models. Each one is involved in a different stage of data analysis supplementing each other. This way, the system can detect which user is indeed a fraud or not. Once there are some specific actions noticed, the account may be limited or even banned.