A framework for understanding this research is sketched in Figure 25.1. First, we
demonstrate the extent of the problem by modeling efficient attacks, attacks that can
with relatively low cost produce a large impact on system output. This enables us
to understand the shape of the impact curve for efficient attacks. Research on detection attempts to identify groups of profiles that make up an attack and to eliminate
them from the database. Attacks that are not efficient are more difficult to detect,
but because they are inefficient, must be very large to have an impact. A large influx
of ratings for a particular item is easy to detect with standard system monitoring
procedures. Research on detection therefore focuses on how to detect efficient at-
tacks and variants of them, seeking to increase the size of the “detectable” boxes in
the diagram, and thereby limiting the impact that an attacker can have. At the same
time, researchers have studied a number of algorithms that are intended to be ro-
bust against attack, having lower impact curves relative to efficient attacks. With the
combination of these techniques, researchers have sought, not to eliminate attacks,
but to control their impact to the point where they are no longer cost-effective.
This chapter looks at each of these points in turn. In Section 25.3, we look at
research that aims to identify the most efficient and practical attacks against collaborative recommender systems, establishing the shape of the impact curve suggested
above. Section 25.5 looks at the problem of detection: in particular, the left-most
shaded area for detecting efficient attacks. Lastly, in Section 25.6, we examine attempts to reduce the impact of attacks through robust algorithms