FL-MBS: Federated Learning and Mantissa Bits Steganography for Resource Constrained Autonomous Uncrewed Vehicles

In the modern Internet of Battlefield Things (IoBT), autonomous uncrewed vehicles (AUVs) transmit critical messages to military bases while training machine learning models using federated learning (FL). Due to the resource-constrained nature of IoBT, it is imperative to design resource-efficient systems. Therefore, to optimize communication and resource usage, we propose FL-based mantissa bits steganography (FL-MBS), enabling simultaneous FL training and message transmission over a single communication link. Specifically, we encode the AUV messages into an FL weights matrix that is transmitted to the military base. We demonstrate FL's efficacy in IoBT by comparing it to centralized learning on a custom IoBT dataset for binary classification. Using the same dataset, we evaluate the FL-MBS framework, analyze tradeoffs between FL performance and bits per parameter across four perturbation scenarios, and propose a self-recovery method to maximize message embedding capacity (MEC) while maintaining FL model performance. Our approach achieves a MEC of 240.4 KB, which is further maximized to 368.61 KB using the self-recovery method. In addition, we demonstrate the defensive capabilities of our proposed approach against various FL attacks, further enhancing the security of battlefield FL data and transmitted messages in IoBT.