Indirect trust computation based on recommendations form an important component in trust-based access control models for pervasive environment. It can provide the service provider the confidence to interact with unknown service requesters. However, recommendation-based indirect trust computation is vulnerable to various types of attacks. This paper proposes a defense mechanism for filtering out dishonest recommendations based on a measure of dissimilarity function between the two subsets. A subset of recommendations with the highest measure of dissimilarity is considered as a set of dishonest recommendations. To analyze the effectiveness of the proposed approach, we have simulated three inherent attack scenarios for recommendation models (bad mouthing, ballot stuffing, and random opinion attack). The simulation results show that the proposed approach can effectively filter out the dishonest recommendations based on the majority rule. A comparison between the exiting schemes and our proposed approach is also given.