This talk is based on results of R&D project aimed to build a solution for user behavior security analytics. I will describe various methods and ideas for anomaly detection solutions built to understand user behavior trends and find abnormal activity using state-of-the-art neural networks.
The talk covers things like:
Empowering a feature selection process with clustering algorithms
Checking the quality of data with a serial correlation algorithm
Implementing a behavioral whitelisting with scoring analysis
Tuning a scoring engine with frequency analysis
Advancing scoring engine results with prefix coding
Predicting user actions with recurrent neural networks
Adaptive threshold for RNN predictions
Generating synthetic but realistic datasets
Peer group analysis and other interesting ideas.