Commercial drones have become much popular in recent years. Their low cost and high capabilities have rendered drones feasible for different applications. These include weather observations, traffic monitoring, inspection of buildings, fire detection, delivery of goods, agriculture and security. However, the abuse of the advancements of drone capabilities can lead to security and public safety issues. Such abuses include the transfer of illegal or dangerous goods, assaults and terrorizing actions, espionage or spying. In such a context, can the behavior of a quadcopter be determined from observations? Can those observed behaviours be utilized for training a machine learning model for classifying future comportment? In this work, we look at three pieces of information that we can predict about a quadcopter or a group of quadcopters, leveraging behavior observations. First, we try to predict the formation a group of drones intend to make while in transition by training a machine learning model, based on Softmax regression, on navigational data collected for this purpose. Second, we train an Long Short-Term Memory (LSTM) neural network on Mel Frequncy CepstralCoefficients (MFCCs), which are audio features extracted from an acoustic signal, to predict the weight of the payload of a quadcopter. Furthermore, in our third task, we identify whether a drone pilot is a human or an autopilot using a deep neural network trained on features processed from collected navigational data. In each of the three tasks, we evaluate features extracted from the data, then build and evaluate different models. The best-performing models in each of the three tasks are then compared to three different dummy classifiers. This comparison is made by performing statistical significance tests demonstrating that the difference in performance is not just a result of a statistical fluke. The three dummy classifiers do not learn any patterns. However, they are based on different classifying strategies, such as always using the most frequent label, generates predictions by respecting the training set's class distribution, or generates predictions uniformly at random.