Recent works suggest to model the reputation of the agents in order to evaluate the sincerity of an agent. The very first step of this mechanism is the detection of fraud and lie. However, most of the works concerning reputation consider this detection as implicit either because it is obvious in their specific application or because they assume that this detection is already made. In the present paper, we propose a definition of the concepts of lie, deception and fraud. From these definitions, we also propose a decentralized detection of lie and deception. Finally, this detection mechanism is used to maintain a representation of the agents' reputation.