A recommendation-based centralized trust model was proposed to deal with three kinds of common attacks including self-promoting attack, slandering attack and whitewashing attack. The model calculated the similarity of consumption custom between buyers and the timeliness of the buyers' evaluation opinions to determine the recommendation trust value, and obtained the direct trust value on the basis of the buyer's own trading experience. According to the direct and recommendation trust value, composite trust value could be calculated. Simulations were given by injecting the three kinds of attacks into real transaction data. The results prove that, compared with the existing reputation management models, the proposed model can effectively prevent attacks, and provides more accurate reputation information of sellers and goods to consumers.