Online systems present users with a vast amount of information and product/service offerings. In order to deal with this type of information overload, many top online websites use recommender systems to provide users with a more personalized, and less daunting, experience by providing recommendations on what to consume. As with many online systems, there is a potential for malicious users to "game the system" for personal benefit or pleasure. This constitutes an "attack" on recommender systems and usually consists of having malicious users enter a number of fake ratings or reviews in order to promote or disparage an item for personal gain, or just to disrupt the system's recommendations. The problem with attacks on recommender systems is that they bias the underlying data and cause the system to deliver erroneous or misleading recommendations to online users. This can cause users to lose trust in the system and either (1) do business elsewhere, negatively impacting the sales of the attacked product/service provider, or (2) purchase the product/service only to find out that it does not meet their needs, negatively impacting user satisfaction with the online recommender. This dissertation extends the body of knowledge of attack models by designing novel attacks based on the "influence" characteristics of users within a recommender system. For collaborative recommenders, i.e., those based on ratings similarities between users, a social graph describing the relationships between users can be defined. And from Social Network Analysis, we know that central, or "power", users are those that wield influence over other users. In the recommender system context, the term "power user" denotes users who have considerable influence over the recommendations presented to other users. Thus, this research formulates the power user attack model and evaluates the accuracy and robustness impacts that power user attacks can have on recommender systems. The results of this research indicate that (1) Power user attacks can have significant impact on the predictions generated by popular collaborative recommender algorithms across the movie and music domains tested, i.e., these attacks can efficiently and effectively bias the recommender predictions as measured by accuracy and robustness metrics, (2) Synthetic power user profiles generated from the In-Degree and Number of Ratings power user selection methods result in effective power user attacks, (3) Due to its low cost of attack, the simple Number of Ratings method is the most efficient and effective approach for selecting and generating power user profiles, and (4) Reducing the influence of power users is a more effective and less impactful mitigation strategy than completely removing power users from the dataset for user-based recommenders in the movie and music domains tested.