People are highly dependent on online social networks (OSNs) which have attracted the interest of cyber criminals for carrying out a number of malicious activities. An entire industry of black-market services has emerged which offers fake accounts based services for sale. We, therefore, in our work, focus on detecting fake accounts on a very popular (and difficult for data collection) online social network, Facebook. Key contributions of our work are as follows. The first contribution has been collection of data related to real and fake accounts on Facebook. Due to strict privacy settings and ever evolving API of Facebook with each version adding more restrictions, collecting user accounts data became a major challenge. Our second contribution is the use of user-feed information on Facebook to understand user profile activity and identifying an extensive set of 17 features which play a key role in discriminating fake users on Facebook with real users. Third contribution is the use these features and identifying the key machine learning based classifiers who perform well in detection task out of a total of 12 classifiers employed. Fourth contribution is the identifying which type of activities (like, comment, tagging, sharing, etc) contribute the most in fake user detection. Results exhibit classification accuracy of 79% among the best performing classifiers. In terms of activities, likes and comments contribute well towards detection task. Although the accuracy is not very high, however, our work forms a baseline for further improvement. Our results indicate that many fake users are classified as real suggesting clearly that fake accounts are mimicking real user behavior to evade detection mechanisms. Our work concludes by enlisting a number of future course of actions that can be undertaken.