Interestingly, Figure 25.3 (right) shows that the largest hit ratios were achieved
by the bandwagon attack, indicating that prediction shift does not necessarily translate directly into top N recommendation performance. This result is particularly
encouraging from the attacker’s perspective, given that the required knowledge to
implement such attacks is low. Note that all attacks significantly outperform the
pre-attack hit ratio results (indicated by “baseline” in the figure).
The item-based algorithm was shown in [12] to be relatively robust against the
average attack. The segment attack, introduced in [19], was specifically crafted as
a limited-knowledge attack for the item-based algorithm. It aims to increase the
column-by-column similarity of the target item with the users preferred items. If the