Blockchain systems have been successful in discerning truthful information from interagent interaction amidst possible attackers or conflicts, which is crucial for the completion of nontrivial tasks in distributed networking. However, the state-of-the-art blockchain protocols are limited to resource-rich applications where reliably connected nodes within the network are equipped with significant computing power to run lottery-based Proof-of-Work (PoW) consensus. The purpose of this work is to address these challenges for implementation in a severely resource-constrained distributed network with Internet of Things (IoT) devices. The contribution of this work is a novel lightweight alternative, called weight-based reputation (WBR) scheme, to classify new transactions via modeling blockchain decisions as a distributed machine-learning task. WBR identifies network nodes that are willing to cooperate toward securing ground truth, showing robustness to adversarial subnetworks that are greater than 50% and reducing collaboration error by 50% compared to other similar schemes. This two-step approach of reputation plus transaction classification for generating blockchain data is treated as a novel method of preventing fraud and double-spending attacks in blockchain networks. To capture adversary influence, a Bayesian game is formulated and implemented to show superior performance to the state-of-the-art along with resource consumption metrics.

Decision-making and consensus in traditional blockchain protocols is formulated as a repeated Bernoulli trial that solves a computationally-intense lottery puzzle, called Proof-of-Work (PoW) in Bitcoin. This approach has shown robustness through practice, but does not scale with increasing network size and generation of new transactions. Resource constrained Internet of Things (IoT) networks are incompatible with full computation of schemes like Bitcoin's PoW. Our effort proposes a first step towards an alternative consensus using machine learning-based decision-making with prediction of fraud transactions to alleviate need for intense computation. To improve base approval probabilities for fraud detection in an ideal security setting, Vector GAN (VecGAN) is proposed to augment blockchain data in classifier training, which combines error-driven learning with Bayesian estimation to alleviate calculations. This two-step approach with augmentation and classification on new transactions is proposed as a novel approach to blockchain decision-making. Experimental prediction accuracy using VecGAN improved up to 3% on simplistic classifiers compared to other state-of-the-art augmentation techniques. Resource consumption in a realistic blockchain setting was reduced while improving block throughput by 50% compared to PoW. Future work will explore Sybil-spam defensive measures for realistic protocol implementation with this approach.