This study examines how artificial intelligence (AI) can be incorporated into the Internet of Things (IoT), with a particular focus on using the NVIDIA Jetson Nano as a specialized AI platform. The aim is to boost the functional capabilities of IoT devices, allowing them to independently carry out sophisticated data analysis and decision-making tasks to heighten their efficiency and effectiveness. The paper begins by outlining the motivation for this investigation, then proceeds with a literature review to evaluate the prospective applications, challenges, and benefits of AI in IoT environments. It explores in detail the technical attributes of the NVIDIA Jetson Nano, including its compact design, robust computing power, and the availability of software libraries that facilitate the deployment of AI models on devices with limited resources. A prototype IoT device equipped with the NVIDIA Jetson Nano was developed at Tien Giang University. This device is capable of capturing video and executing AI algorithms for real-time analysis, thereby deriving crucial data for IoT decision-making processes. The outcomes and structure of this prototype exhibit the potential for wider application, particularly in facial recognition technologies. The discussion on this subject extends beyond the confines of this document, indicating ongoing developments and enhancements.

This study examines improvements in stegano-graphic models by redesigning three main components: the encoder, decoder, and evaluator, each optimized for better performance. The encoder is designed to embed data effectively with minimal quality, using popular activation function like CeLU. The decoder focuses on accurately retrieving embedded data, essential when images are processed under various conditions. The evaluator uses new functions to accurately check how well data is recovered, promoting continuous improvements. Experimental tests conducted on the Div2K and COCO datasets revealed varied results. The Basic model managed to embed a lot of data (BPP of 4.160), but this reduced image quality significantly, as shown by a PSNR of 30.026 and SSIM of 0.824. In contrast, the Residual model preserved high image quality, with a PSNR of 37.113 and SSIM over 0.9, but embedded less data (BPP of 4.127). The Dense model struck a good balance, embedding a fair amount of data (BPP of 4.512) while maintaining decent image quality (PSNR of 37.089 and SSIM of 0.907). These outcomes demonstrate the potential benefits of making targeted changes to model architecture, suggesting great promise for improving security features in digital communication tools.