In this literature, we concentrate on investigating a learning-based resilient predictive control framework using variable-step event-triggered mechanism, which aims to avoid unnecessary events and enhance the system robustness subject to actuator false data injection (FDI) attacks. To be more precise, to improve the robust performance of the controlled system under both actuator attacks and parametric uncertainties, a learning-based robust model predictive control (MPC) architecture is developed. In this control architecture, an online learning strategy is incorporated into a neural network weight update policy, which can provide a reinforced structure and accelerate the learning process. Meanwhile, in order to circumvent the unnecessary triggering and commutation behavior, a tentative verification of a triggering condition and a delayed triggering with a variable-step waiting horizon are embedded into the suggested event-triggered mechanism. The main feature of our development is that it not only enhances the control property under the actuator FDI attacks, but also attenuates the inherent issues of unnecessary switching losses and parametric uncertainties affecting the system, opening a wide research field for resilient finite control-set MPC. Finally, we highlight its advantages with a case study.