<p>Selfish mining is a mining strategy that aims to gain disproportionate rewards by manipulating the block propagation and synchronization mechanisms in a blockchain network. The primary method involves delaying the publication of mined blocks to control the progress of the blockchain, allowing selfish miners to receive more rewards. However, existing research on selfish mining mainly focuses on evaluating the ability of Proof-of-Work blockchains to resist selfish mining by adjusting consensus and network parameters, using different models to calculate the rewards from selfish mining, and evaluating how these parameters impact the Proof-of-Work system’s resistance to attacks. The current detection methods also have limitations. For example, these studies perform detection on simulated blockchain networks, which allows them to access data that is not available in actual blockchain networks, making such methods unrealistic. Moreover, they do not take into account such characteristics as transaction fee and generation time. To fill the gap, we propose SeMi_Detector, a new method based on a multilayer perceptron to classify block data. We crawled real Bitcoin block data and obtained forking information from DSN, filtering and aggregating the data to extract relevant features. First, we selected a subset of data and labeled it, then trained the labeled data using a multilayer perceptron. Afterward, we applied pseudo-labeling to all the data. Finally, we used both the labeled and pseudo-labeled data for training and detection. To address the issue of class imbalance between anomalous and normal samples, we introduced a weight balancing strategy. Experimental results show that SeMi_Detector successfully detects blocks generated by the selfish mining strategy with the precision of 93.75% and recall of 89.29%, outperforming other methods. Overall, the SeMi_Detector is a novel and effective approach for selfish mining detection that can be applied in real blockchain networks and demonstrates strong accuracy.</p>