This paper introduces a unified learning framework for Byzantine-resilient spectrum sensing and secure transmission in intelligent reflecting surface (IRS)-assisted networks under channel state information (CSI) uncertainty. The sensing module employs robust Bayesian belief updates with adversary-resistant aggregation and consensus, guaranteeing reliable primary user (PU) detection even when a bounded fraction of users are malicious. Based on the sensing outcome, the transmission module formulates the downlink design as a sum mean-squared error (MSE) minimization problem under transmit-power and signal-leakage constraints, jointly optimizing the base station (BS) precoder, IRS configuration, and user equalizers. For partial or known CSI, we develop a lightweight alternating optimization algorithm with provable sublinear convergence. For unknown CSI, we integrate constrained Bayesian optimization (BO) within a geometry-aware, low-dimensional latent space. Simulations demonstrate that the proposed framework achieves a higher probability of detection at a fixed false-alarm rate under adversarial attacks compared to state-of-the-art schemes. It also yields substantial reductions in user MSE, strong suppression of eavesdropper signal power, and fast convergence. This work provides a practical, resilient solution for coherent sensing–communication coordination in emerging sixth-generation (6G) applications such as vehicular, unmanned aerial vehicle (UAV), and Internet of Things (IoT) networks.