Recently, the costs of training deep learning models have dramatically increased. Volunteer Computing (VC) enables cost-effective gathering of large amounts of idle computing resources, but has been little applied to deep learning. Volunteer Deep Learning (VDL) applies VC to train deep learning models while keeping costs low. In this paper we review current VDL concepts, projects, challenges and opportunities. We analyze the VDL security challenges and countermeasures. We propose guidelines towards a VDL framework implementation based on decentralized federated learning.