For increasing serious phishing attacks, machine-learning method is proposed to detect phishing webs. Firstly, sensitive features are extracted from the URL, then, using AdaBoost algorithm to get the trained classifier, and then the classifier is used to detect unknown URLs. Finally, considering of non-equilibrium problems of AdaBoost, the paper puts forward the improved learning algorithm called AdaCostBoost, which contains computation of cost factors. According to the experiment result, the proposed phishing detection method has better detection performance.