Accurately Detect Bots in Blockchain System

Companies care about bots a lot! While good bots can be used to automatically test the system’s effectiveness, bad bots can “disguise” as normal business activities and block consumer usage and swipe their data, causing an average of $6 million in transaction volume per day! However, bot behaviors can vary so much and be so similar to human account behaviors, making it very difficult for companies to distinguish true business transactions and manipulated ones. Anchain Group 2 decided to team up with Anchain, a Silicon Valley-based blockchain security company to use their data on DApp gameplay in blockchain system to develop ways to clearly and accurately identify human and bot accounts using feature engineering and machine modeling methods.

The challenges are to find the best features that can distinguish human and bots, to develop the most appropriate model to classify human/bot behavior, and to visualize our findings and models to be used by the companies. Anchain Group 2 is able to address 12 features that include aspects of Activity, Transaction, and Account Type to indicate accounts’ behavior patterns. By combining manual labeling check and undersampling of Logistic Regression Modeling, the team is able to accurately classify humans and bots from existing accounts and is able to move on to more complex models, such as Boosting, Random Forest, and CART. All three models reached 100% accuracy.

Finally, Anchain Group 2 applied feature importance of Random Forest model to show the list of important features, visualized number of human/bots and behavior of them by T-distributed Stochastic Neighbor Embedding, and developed a fulfilled search engine and UI to enable users to check the dataset’s account type (human/bot) and behavior indicators.

Read more at https://anchain-ai.weebly.com/!

Built by: Anderson Lam, Joshua Nuesca, Simone Ong, Matt Wong, Michelle Xiao