Deep Anomaly Detection for Fraud Detection in Cyber Security
Tech, data & sensors:
Time-series, Network signal
Deep Anomaly Detection, Deep Generative Models
The mission is to protect the internet by verifying the humanity of every online transaction working with the web’s largest platform which serve trillions of monthly decisions. This kind of technology protects digital advertisers, publishers, ad tech, and enterprise businesses globally by detecting and blocking bots. Traditionally, tackling this kind of problem companies are based on in-depth technical evidence regarding the nature of each request. In contrast, we applied machine learning techniques to leverage hidden information beyond human perception.
A continuous training solution was implemented employing state-of-the-art DL to detect invalid (i.e. non human) requests based on historical evidence. The power of deep anomaly detection modeling was utilized, employing deep neural networks for modeling the normal behaviour of a user. In this way we were able to identify abnormal user activities by scoring different users and classifying them as fraudulent or not. Multiple deep generative models have been tried such as simple LSTM Autoencoder (AE), conditional, denoising, and compositional Autoencoder, RandNet and generative adversarial networks (GANs).
This work greatly improved bot detection performance, improvement which is crucial for a business that wants to disrupt one of the biggest cyber crime activities. In addition, the specific solution opened new directions and opportunities for further tracking invalid activity by employing higher level features and better behavioral modeling.