25.2.1 An Example Attack
To illustrate the basic idea of a profile injection attack, consider the simplified recommender system database that is presented in Figure 25.2. In this example, the
objective is to demote the recommendations that are made for item 7 (i.e. a product
nuke attack), and a number of attack profiles (users i through m) have been inserted
into the system to target this item.
In particular, consider the binary recommendation problem in which the task is
to predict whether or not user h likes item 7. In the first instance, let the attack
profiles be ignored and consider only the authentic profiles (users a through g) as
possible neighbours for the target user, h. Regardless of the specific recommendation
algorithm used, presumably the algorithm would determine that users a and f have
similar tastes to the active user, and since both of these users like item 7, a positive
recommendation for the item follows.
When the attack profiles are also considered as possible neighbours, the situation
is significantly altered. Several of these attack profiles are also similar to user h, and,
since all of these profiles rate item 7 poorly, the system is now likely to recommend
a negative rating for the item. Thus, the objective of the attack is realised. The next
section discusses how these attack profiles must be crafted to work well in a realistic
setting.