Recent works have highlighted how misinformation is plaguing our online social networks. Numerous algorithms on automated misinformation detection are centered around deep learning~(DL) which requires large data for training. However, privacy and ethical concerns reduce data sharing by stakeholders, impeding data-driven misinformation detection. Current data encryption techniques providing privacy guarantees cannot be naively extended to text inference with DL models, mainly due to the errors induced by stacked encrypted operations and polynomial approximations of the otherwise encryption-incompatible non-polynomial operations. In this paper, we show, formally and empirically, the effectiveness of (1) $L_2$ regularized training to reduce the overall error induced by approximate polynomial activations, and (2) sigmoid activation to regulate the error accumulated due to cascaded operations over encrypted data. We assume a federated learning-encrypted inference~(FL-EI) setup for text-based misinformation detection as a (secure and privacy-aware cloud) service, where classifiers are securely trained in FL framework and inference is performed on homomorphically encrypted data. We evaluate three architectures—Logistic Regression~(LR), Multilayer Perceptron~(MLP), and Self-Attention Network~(SAN)—on two public text-misinformation datasets with some interesting results, for example, by simply replacing ReLU activation with sigmoid, we were able to reduce the output error by $1750\times$ in the best case to $43.75\times$ in the worst case.