Digital watermarking is essential for protecting intellectual property and verifying content authenticity. While recent deep learning-based methods have advanced robustness and visual quality, they often rely on large, computationally intensive models. We present EnhanceGuard, a lightweight yet robust watermarking framework with only 1.39 million parameters, just 28% of the parameter count of EditGuard (4.96M), achieving an effective balance between computational efficiency and robustness. EnhanceGuard adopts a simplified UNet-inspired architecture with three down-up sampling stages, enhanced by residual and attention modules for effective feature fusion. By replacing InstanceNorm with BatchNorm2D, our model improves generalization and training stability, as confirmed by extensive ablation studies. Despite its compact design, EnhanceGuard achieves up to 99.93% bit accuracy under common distortions (Gaussian noise, JPEG compression, Poisson noise) and remains highly robust against regeneration and adversarial attacks. It also attains an average PSNR of 38.56 dB, reflecting high visual fidelity of the watermarked images. These results highlight EnhanceGuard as a potential solution for real-world watermarking applications, especially in resource-constrained environments. In addition, by embedding metadata about key objects and their locations, EnhanceGuard can be extended to not only certify authenticity but also localize tampered regions, providing examiners with direct evidence of where an image has been modified.