Federated Learning (FL) is a growing technology that enables training of Deep Learning models on private data. Many FL enhancements have been proposed, notably for better security and privacy. Current architectures and frameworks focus on specific sets of enhancements with little extensibility and do not support composition of enhancements. In this paper, we introduce Ti-skol, an architecture and framework that supports composition of security and privacy countermeasures, including countermeasure incompatibilities. Ti-skol also enables modular management of FL enhancements beyond security, being compatible with most enhancements. Ti-skol is promising to assess the cost of countermeasures, individually or in combination. We evaluate our framework on a use-case of Volunteer Deep Learning -- applying Volunteer Computing to reduce the cost of large model training by harnessing idle resources of single machines into the required massive distributed computing power. Experimental results show that Ti-skol is scalable as the network size increases. While adding security countermeasures such as Byzantine protections or secure aggregation substantially increase computing overheads, they do not change their order of magnitude, individually or in combination. This tends to show the practicality of the Ti-skol framework for on-demand FL security.