Phishing-Test-All-Datasets

Phishing scams on the Ethereum network are in-creasingly severe, causing significant financial losses to users.In recent years, researchers have proposed various methods todetect phishing scams on Ethereum. However, the real-worldeffectiveness of these models is uncertain, as experimental dataoften diverges significantly from actual data distributions, withsome datasets being filtered or preprocessed, and the character-istics of phishing nodes on Ethereum evolving over time. In thisstudy, we conduct the first systematic analysis of effectiveness ofdifferent phishing detection methods on Ethereum. We collectedand thoroughly examined research papers that focus on phishingdetection on Ethereum. We aggregated all the datasets disclosedin these papers and combined them with additional data wegathered from Ethereum, creating a new and comprehensivedataset that is the largest of its kind.We systematically evaluate ten representative phishing detec-tion models using a unified dataset segmented by different dataratios and data time. The experiments revealed that scalabilityis a significant challenge, with most models, except Trans2Vec,struggling to handle larger, imbalanced datasets, particularly at a1:50 ratio. And as the positive-to-negative sample ratio increases,model performance generally declines. Additionally, the time-series dataset analysis shows that models such as BERT4ETHmaintain stability across time, while others, such as EPD, exhibitsignificant performance degradation when tested on more recentdata. These findings underscore the need for adaptable androbust detection methods that can handle both evolving phishingpatterns and imbalanced datasets