Description: 
Publicly-accessible adaptive systems such as recommender systems present a security problem. Attackers, who cannot be readily distinguished from ordinary users, may introduce biased data in an attempt to force the system to “adapt' in a manner advantageous to them. The ability to understand, identify, and defeat such “bias injection' attacks will have significant implications for a variety of adaptive information systems that rely on users' input for learning user or group profiles. Many such systems have open components through which a malicious user or an automated agent can affect the overall system behavior.

Recent research has begun to examine the vulnerabilities and robustness of different recommendation techniques, such as collaborative filtering, in the face of bias injection attacks. In this presentation, I will outline some of the major issues in building secure recommender systems, concentrating in particular on the modeling of attacks, their impact on various recommendation algorithms, and methods for automatic detection of attack profiles. I will introduce several new attack models not previously studied and present simulation-based evaluation results to show which attack models are most successful against common recommendation techniques. The evaluation criteria will consider both the overall impact on the ability of the system to make predictions and generate recommendations, as well as the degree of knowledge about the system required by the attacker to mount a realistic and successful attack. Our study, to date, shows that standard collaborative filtering algorithms are highly vulnerable to specific attack models, but that hybrid algorithms, which integrate semantic knowledge about items with user profiles, may provide a higher degree of robustness.

Runtime:01:11:59

Rating:TV-G