This project will evaluate vulnerabilities in relatively unstudied model-based recommendation systems in which recommendations are based on a model that relates ratings on one item to ratings on other items. Recommendation systems are a means of reducing "information overload" by filtering a potentially overwhelming number of options (such as all the products available from a seller) to identify those calculated to be of greatest interest. This project extends research on collaborative recommendation systems, which base recommendations for an individual on the preferences expressed by other people, by investigating the problem of malicious manipulation of these systems, for example, by an attacker attempting to influence the outcome with biased or faked rating profiles. Research suggests that a specific model-based systems exhibit much more resistant to recommendation attacks than memory-based systems in which recommendations are based on the principle of finding similar users or similar items. Moreover, this research will identify any previously unknown attack methods that might be specifically effective against model-based recommendation systems.

Agency
National Science Foundation (NSF)
Institute
Division of Computer and Network Systems (CNS)
Type
Standard Grant (Standard)
Application #
0716827
Program Officer
Carl Landwehr
Project Start
Project End
Budget Start
2007-09-01
Budget End
2010-08-31
Support Year
Fiscal Year
2007
Total Cost
$228,374
Indirect Cost
Name
University of California Davis
Department
Type
DUNS #
City
Davis
State
CA
Country
United States
Zip Code
95618