Physical-layer secret key generation (PSKG) has emerged as a promising technique for enhancing wireless security in Internet of Things (IoT) networks by exploiting the reciprocity of uplink and downlink channels. However, in time-division duplex (TDD) systems, hardware impairments and channel noise disrupt channel reciprocity, degrading key generation performance. To overcome these challenges, this paper introduces a deep learning–enhanced PSKG framework that effectively mitigates channel discrepancies and improves key generation reliability under imperfect channel state information (CSI). Specifically, a two-dimensional convolutional neural network–based autoencoder (2D CNN–AE) with a spatial self-attention (SSA) mechanism is developed to efficiently extract and learn channel reciprocity features in time-division duplex (TDD)-based fifth-generation (5G) networks. Additionally, a quantile-based quantization scheme is proposed to enhance key randomness and entropy, thereby strengthening security and resilience against potential threats. To facilitate comprehensive performance evaluation, a wiretap channel dataset is generated in accordance with 5G standards, encompassing diverse propagation conditions and including both legitimate users and an eavesdropper. Extensive simulation results demonstrate that the proposed 2D CNN–AE–SSA-based PSKG framework significantly reduces the key disagreement ratio (KDR), enhances key randomness, and maintains low computational complexity. These findings establish the proposed method as a robust and practical solution for securing wireless communications in resource-constrained IoT environments.