A study of attacks on collaborative filter
MetadataShow full item record
Collaborative filtering is a widely used technique to make classifications by using distributed feedback from all users. Recently, collaborative filtering has been proposed and used to detect spam messages. Compared to individual spam filters, collaborative spam filtering has the advantage of potentially accessing large datasets and the effect of crowd sourcing. However, the benefit of the collaboration also comes along with vulnerabilities of false collaborative information. In this thesis, I studied the effect of various possible attacks on a collaborative spam filtering system. We built a platform to simulate a collaborative spam filtering system, and use this system to answer questions such as what attacks cause the most damage and the cost of launching such attacks. The result of this study suggests that collaborative spam filtering is vulnerable to attacks from false collaborators and we identified the most damaging strategy that should be addressed first by the defense of collaborative spam filtering.