Abstract: The privacy protection and credible authentication of employees of energy companies are the key to safe production. Federated learning can construct a user authentication model while protecting the privacy of the data itself. However, in the process of federated learning, there are defects such as low data collection efficiency, nodes susceptible to poisoning attacks, and lack of practicability. To this end, this paper proposes a new mobile user authentication model optimization method that combines smart contract private chain and federated learning. By combining discrete wavelet decomposition of time series data and long-short term memory network, this method could increase the data collection rate, reduce the probability of a trained model being bypassed by an attacker and improve the security of the model. The experimental results on the real environment data set of 1513 people show that the proposed method can effectively resist poisoning attacks and achieve a relatively satisfactory accuracy.