A decentralized technique of machine learning called federated learning improves security by enabling local training. No research has yet to compare various deep learning architectures with federated learning. This work integrates federated learning with deep learning for the detection of malware. It proposes a comparative federated learning analysis study of different neural architectures including Artificial Neural Network, Gated Recurrent Unit, Long Short-Term Memory, and Convolutional Neural Network. Both performance and security are analyzed. The evaluation is conducted on Malware Dataset and AndroMD Dataset. The security of all models are evaluated against model inversion attack. For both datasets, FL_ANN is the fastest model while FL_LSTM is the slowest model. The highest performance metrics are achieved by FL_CNN on Malware Dataset and FL_LSTM on AndroMD Dataset. The FL_ANN is the most robust model using Malware Dataset with an average MSE of 1.91 while FL_LSTM is the best resistance model using AndroMD Dataset with an average MSE of 1.44.