Electromagnetic information leakage has emerged as a critical side-channel threat, enabling adversaries to reconstruct sensitive display content by capturing unintended emissions from electronic devices. This phenomenon, known as TEMPEST, poses heightened risks in the Internet of Things (IoT) era, where numerous connected devices with embedded displays are deployed across industrial, financial, and consumer domains. Deep learning has improved TEMPEST image reconstruction, but most existing methods rely on synthetic datasets with simple characters. We propose an attention-enhanced DRUNet that integrates an efficient multi-scale attention (EMA) module into skip connections to improve contextual representation and suppress noise. Experiments on a realistic TEMPEST dataset show consistent gains over state-of-the-art baselines in peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and character error rate (CER), demonstrating the effectiveness of the proposed approach for practical TEMPEST reconstruction and IoT security.