Efficient bot detection is a crucial security matter and has been widely explored in the past years. Recent approaches supplant flow-based detection techniques and exploit graph-based features, incurring however in scalability issues in terms of time and space complexity. Bots exhibit specific communication patterns: they use particular protocols, contact specific domains, hence can be identified by analyzing their communication with the outside. To simplify the communication graph, we look at frequency distributions of protocol attributes capturing the specificity of botnets behaviour. In this paper, we propose a bot detection technique named BotFP, for BotFinger-Printing, which acts by (i) characterizing hosts behaviour with attribute frequency distribution signatures, (ii) learning behaviour of benign hosts and bots through a clustering technique, and (iii) classifying new hosts based on distances to labelled clusters. We validate our solution on the CTU-13 dataset, which contains 13 scenarios of bot infections, connecting to a Command-and-Control (C&C) channel and launching malicious actions such as port scanning or Denial-of-Service (DDoS) attacks. Our approach applies to various bot activities and network topologies. The approach is lightweight, can handle large amounts of data, and shows better accuracy than state-of-the-art techniques.