This paper researches mean-square bounded consensus for a nonlinear multi-agent system subjected to randomly occurring deception attack, process and measurement noises. Considering the measurement tampered by the attacker, an estimator is presented to obtain relative accurate state estimation, where the gain is acquired by a recursive algorithm. On this basis, a type of centralized controller is designed combined with cloud computing system. Moreover, from perspective of the defender, a detector is proposed at the side of agent to detect whether the current actuator input is attacked. Using linear matrix inequality, sufficient conditions are given for achieving mean-square bounded consensus and an upper boundary is derived. Finally, validity of the proposed method is illustrated via two simulation examples.