Smart Ponzi Scheme Detection using Federated Learning

Decentralized applications (DApp) have opened up new heights for the use of blockchain and cryptocurrencies, but users with poor intentions utilize DApp websites to promote smart Ponzi schemes, causing huge losses to inexperienced investors. To alert potential investors, DApp websites tend to mark high-risk DApps uploaded by users. So far, previous works have only trained the classifiers of smart Ponzi schemes in a unilateral way. However, smart contract dataset is very skewed, there are much fewer smart Ponzi schemes than non-Ponzi schemes. Besides, due to data security and privacy, different DApp websites are generally not allowed to share their users’ data and information. As a result, it is difficult for a certain DApp website to learn the patterns of smart Ponzi schemes and detect them alone.This paper proposes a novel smart Ponzi scheme detection framework named SPSD-FL (Smart Ponzi Scheme Detection in Federated Learning) for the first time to address these problems. Instead of unilateral training, our method fulfills a secure aggregation of local gradient histograms by a horizontal federated learning based on XGBoost algorithm, in which we leverage a data augmentation method and labeled smart Ponzi schemes’ code features, dapp submitters’ information, and investors’ information to implement the collaborative training of the detection model between different DApp websites without sharing the original training data and protecting the user’s sensitive information. Experiments conducted on the real dataset show that our method can be trained substantially more data-efficiently and the SPSD-FL model achieves an average F-score of 96.55%, which is better than the unilateral training model and no less than that of the centralized training model.