Collaborative filtering (CF) algorithms fundamentally depend on similarities between users and/or items to predict individual preferences. There are various binary similarity measures like Kulzinslcy, Sokal-Michener, Yule, and so on to estimate the relation between two binary vectors. Although binary ratings-based CF algorithms are utilized, there remains work to be conducted to compare the performances of binary similarity measures. Moreover, the success of CF systems enormously depend on reliable and truthful data collected from many customers, which can only be achieved if individual users' privacy is protected. In this study, we compare eight binary similarity measures in terms of accuracy while providing top-N recommendations. We scrutinize how such measures perform with privacy-preserving top-N recommendation process. We perform real data-based experiments. Our results show that Dice and Jaccard measures provide the best outcomes.