UAVs enhance IoT communication in dynamic, infrastructure-limited settings but face risks of information leakage and eavesdropping. The majority of existing works on covert communication are based on deterministic models and assume a centralized optimization with global CSI, which is not suitable for dynamic UAV–IoT scenarios. To address these challenges, this paper presents a framework, FL–SCA (federated learning–assisted successive convex approximation), for joint secrecy–covertness optimization in UAV-enabled IoT communications. In this context, the distributed IoT devices are trained using the developed FL model to estimate the warden’s detection error probability, while keeping their data private to safeguard their collected data and reduce overused communication resources. The learned detection predictor is used to solve the power-control optimization problem of maximizing the achievable secrecy rate subject to covert communication and transmit-power constraints. Successive Convex Approximation (SCA) is used to solve the non-convex problem and Monte Carlo simulations verify the improvement in terms of secrecy-covertness and resource efficiency. As a result, distributed learning-assisted optimization is applicable to the secure and covert UAV-assisted IoT communication in the dynamic wireless networks.