The growing literature of Federated Learning (FL) has recently inspired Federated Reinforcement Learning (FRL) to encourage multiple agents to federatively build a better decision-making policy without sharing raw trajectories. Despite its promising applications, existing FRL work fails to I) provide theoretical analysis on its convergence; II) account for random system failures and adversarial attacks. Towards this end, we propose the first FRL framework the convergence of which is tolerant to less than half of participating agents being random system failures or adversarial attackers. We prove that the sample efficiency of the proposed framework is guaranteed to scale with the number of agents, accounting for such potential failures or attacks. We empirically verify all theoretical results on various RL benchmarking tasks.