With the rapid advancement of blockchain technology, cryptocurrencies based on blockchain have become a hot topic. However, various issues accompany this development, with phishing scams emerging as a severe financial crime within the blockchain ecosystem, causing significant economic losses to both blockchain platforms and users. In order to address this threat, this essay proposes a phishing scam account identification model based on Bidirectional Long Short-Term Memory Networks (BiLSTM) named BILAM. The model has been validated on the Ethereum platform and has been proven to be effective.This study proposes a novel approach by using transaction records for the first time to construct a time series, and it leverages the BILAM model to learn latent information. Experimental results demonstrate the effectiveness of this method in constructing transaction time series. Moreover, the BILAM model shows excellent performance, with its predictive accuracy significantly surpassing other models, particularly achieving an AUC index of 92.8%.