Secure Personalization: Building Trustworthy Recommender Systems∗

Th l f hi j i l h l bili i f d i d li i i h f f e goa o t s pro ect s to exp ore t e vu nera t es o recommen at on an persona zat on systems n t e ace o li i tt k l t h i f h i th i b t d i th d b hi h tt k b ma c ous a ac s, exp ore ec n ques or en anc ng e r ro us ness, an exam ne me o s y w c a ac s can e i d d ibl d f t d recogn ze an poss y e ea e . Example: Bandwagon attack Profiles use Collaborative Recommendation: Highly Vulnerable . popular items to push a target item C ll b i fil i (CF) d i i l d . o a orat ve ter ng recommen at on s common y use i U ' f t d b fil Impact of attack n e-commerce. sers pre erences are represen e y pro es th t i t f ti f d t S h t id tif After attack Before attack a cons s o ra ngs or pro uc s. uc sys ems en y peer users with similar tastes and extrapolates their ratings as likely 0 8 0.9