A multiagent distributed system consists of a network of heterogeneous peers of different trust evaluation standards. A major concern is how to form a requester’s own trust opinion of an unknown party from multiple recommendations, and how to detect deceptions since recommenders may exaggerate their ratings. This paper presents a novel application of neural networks in deriving personalized trust opinion from heterogeneous recommendations. The experimental results showed that a three-layered neural network converges at an average of 12528 iterations and 93.75% of the estimation errors are less than 5%. More important, the model is adaptive to trust behavior changes and has robust performance when there is high estimation accuracy requirement or when there are deceptive recommendations.