Deep learning classification of bitcoin miners and exploration of upper confidence bound algorithm with less regret for the selection of honest mining.

Bitcoin is the most popular cryptocurrency and it uses proof of work protocol for consensus of all transactions in a block. The blocks are to be appended to the digital ledger, the blockchain. The miners compete for the mining of blocks in the main canonical blockchain. A miner can participate in block mining either individually with his computational power or join a mining pool. Here, the classification of crypto address, whether it belongs to a mining pool or an individual miner, is done with a deep learning Keras framework. The classification accuracy of 99.47% is obtained with 100,000 addresses which is higher than the machine learning random forest classification obtained by Kaggle with 22,000 addresses. The miners in mining pools deploy selfish mining or honest mining to mine a block and get the reward accordingly. In block mining, both honest and selfish miners expose the blocks produced by them. The default protocol of the main canonical blockchain leads to the selection of the longest branch of blocks of the selfish miner, discarding the honest miner's block. To alleviate this, we deploy a reinforcement learning algorithm to choose the block with high upper confidence bound value. This selection explores the branch exposed by honest miners. The algorithm is deployed after the first difficulty adjustment algorithm, where there is more selfish mining activity. Our promising results show that the main blockchain exhibits less regret by selecting the honest miner's branch. [ABSTRACT FROM AUTHOR]