Prediction of Blockchain Transaction Fraud Using a Lightweight Generative Adversarial Network
Charles C. Rawlins, S. Jagannathan, Donald C. Wunsch · 2023
Classic blockchain protocol design is centered around a computationally-intense cryptographic scheme, such as Bitcoin's Proof-of-Work (PoW). Network scalability and efficiency are stifled with the computational resources necessary to approve new transactions, thus rendering PoW unsuitable with limited devices like Internet of Things (IoT). As a first step towards alleviating this, a novel lightweight Generative Adversarial Network (GAN) called Vector GAN (VecGAN) is introduced wherein its weights are tuned through a direct error-driven learning approach with a Bayesian estimator for the selection of random noise matrices, called Bayesian Feedback Alignment (BFA), to augment data for improved fraud prediction. The augmented data is subsequently processed by a classifier for prediction. This combination of VecGAN with a classifier is treated as a novel method of blockchain ledger decision-making to approve ground-truth data. By using the two-step process, prediction accuracy using a classifier was improved up to 8% for real-world datasets. Resource consumption comparison to existing IoT blockchain protocols in a realistic simulation environment is also provided, where lightweight approaches for VecGAN in training and noise modeling reduce computation compared to other techniques.