25.4.5 Attack impact
It is clear from the research summarized above that the memory-based algorithms
that form the core of collaborative recommendation research and practice are highly
vulnerable to manipulation. An attacker with fairly limited knowledge can craft attacks that will make any item appear well liked and promote it into many users’
recommendation lists. The “efficient” attacks that have been developed clearly are
a threat to the stability and usability of collaborative systems and thus we see the
justification for the low-scale / high-impact portion of the theoretical curve shown
in Figure 25.1.
To respond to this threat, researchers have examined two complementary responses. The shaded “detection” areas in Figure 25.1 point towards the first response, which is to detect the profiles that make up an attack and eliminate them.
The second approach is to design algorithms that are less susceptible to the types of
attacks that work well against the classic algorithms.