ABSTRACT Block Withholding (BWH) attack threats are serious for blockchain mining pools, as they target the withholding of valid blocks, affecting the functionality and causing unfair reward distribution. To address this issue, this research introduces a Prevention Method for Block Withholding Attacks (PMBWA), which models the miner's behavior on a geometric manifold with a Tangent Position‐aware Bundle Convolutional Neural Network (TPB‐CNN) and a probabilistic Credit‐Level Classification Algorithm (CLCA), improved by adaptive preprocessing and a behavior‐driven reward/penalty system. Unlike current methods based on the threshold model, game theory, or static reputation systems, PMBWA considers the temporal/directional behavioral dynamics and miner relationship dependencies, allowing for effective detection of adaptive BWH attackers who follow the pattern of honest miners. The potential attackers are further examined through posterior probability clustering and tangent bundle signal filtering, ensuring accurate identification of malicious behavior. Through experimental simulations with different volumes of blocks and malicious computing power, it is shown that PMBWA can efficiently decrease the rewards given to malicious miners and at the same time provide fair compensation to honest miners. The proposed system is able to obtain a precision of 98.9% and a recall of 98.9%, thus improving the security and fairness of the mining pool. The above‐mentioned results show that PMBWA is a dynamic and incentive‐compatible defense system that improves the security of the mining pool while maintaining fairness under adaptive BWH attacks.
Namratha et al. (2026) studied this question.