||Machine learning (ML) techniques can be used in Software Defined Coalitions (SDC) to derive efficient network control policies in an automated manner. However, the applica- tion of ML in coalition networks is challenged by their distributed nature and privacy requirements, where each coalition team can only view and act over a portion of the system, referred by its domain. To address this challenge, we adopt a multi-task learning approach where different coalition partners coordinate to train offline a neural network (NN) model, but the online (run-time) inference can take place in a distributed manner, per domain. The key idea is that the multi-task NN is designed so that the desired output policies of one domain can be computed only with the input information of that domain and some abstract only representation of the other domains. Evaluations for representative scenarios highlight the benefits of the proposed approach.