Aggregation of data from multiple sensor nodes is usually done by simple methods such as averaging or, more sophisticated, iterative filtering methods. However, such aggregation methods are highly vulnerable to malicious attacks where the attacker has knowledge of all sensed values and has ability to alter some of the readings. In this work, we develop and evaluate algorithms that eliminate or minimize the influence of altered readings. The basic idea is to consider altered data as outliers and find algorithms that effectively identify altered data as outliers and remove them. Once the outliers have been removed, use some standard technique to estimate a true value. Thus, the proposed data aggregation algorithm operates in two phases: removal of outliers and computation of an estimated true value from the remaining sensor data. Extensive evaluation of the proposed algorithms shows that they significantly outperform all existing methods.