The invention discloses a method and a system for real-time detection of communication fraud base on suspicious behavior recognition. The method comprises an offline model training stage and a real-time fraud detection stage, wherein an action characteristic recognition model and an action characteristic risk prediction model are established, and a voice call and a content of a short message of a strange call are analyzed to detect abnormal and suspicious behaviors in order to carry out fraud prediction. Through a voice-to-text way, a call content of the voice call of a calling party is turned into text information, and the call content and the content of the short message synchronously utilize a natural language processing method to extract action behavior characteristics and judge whether the action behavior characteristics in dialogue include suspicious behaviors, such as privacy information inquiry and malicious commands. According to the technical scheme for the real-time detection of the communication fraud, the quick and accurate communication fraud prevention detection can be realized, and the possibility of fraud of a user is reduced.