With the rapid development of mobile Internet, phishing attacks are becoming more common on mobile phones. This paper proposes an improved naive Bayes algorithm to detect phishing sites. Firstly, for the purpose of ensuring data integrity in the data collection process, we fill in the missing attribute values through the K-means algorithm to obtain a complete data set. Secondly, for the purpose of eliminating low biased estimation of Bayes algorithm, we appropriately enlarge the probability so as to resolve the underflow problem. Thirdly, for the purpose of avoiding neglecting the relationship between attributes, we weight different attribute values so as to improve the correctness rate of detection. Lastly, for the purpose of resolving the small probability of the occurrence of phishing sites in the actual situation, we adjust the probability ratio of phishing sites and trusted sites so as to further improve the correctness rate of detection. Experiments are deployed on the Android 5.0 mobile phone.The experimental results show that our improved naive Bayes algorithm can effectively detect the phishing attacks on the mobile phone with relatively low time.