<p>Named Data Networking (NDN) improves content delivery efficiency through in-network caching, unlike the traditional IP-based architecture. This capability becomes increasingly important in Internet of Things (IoT) environments, where massive volumes of data are continuously generated. However, selecting appropriate contents for cache replacement remains a major challenge, since storing contents with false or temporary popularity can increase access latency and degrade overall network performance. To address this issue, this study proposes SDMART, (Software-Defined Machine Learning Attention-based Replacement), an intelligent and adaptive cache replacement strategy that integrates NDN with Distributed Software-Defined Networking (DSDN) and enhanced deep federated learning. The proposed model employs Transformer-based neural networks combined with fuzzy logic to optimize cache replacement decisions in NDN-DSDN environments. One of the main contributions of this work is the integration of Federated Learning (FL) into the DSDN architecture, enabling locally collected router data to be collaboratively updated at the controller level without directly transferring raw data to a centralized controller. This mechanism not only reduces computational and communication overhead but also preserves the privacy of local network data. Simulation results demonstrate that the proposed SDMART model significantly improves cache efficiency compared with recent cache replacement approaches, including ISCC 2025 and PeNCache 2025. The proposed strategy increases the cache hit ratio, reduces content retrieval delay, and enhances overall network performance. Specifically, the cache hit ratio improved by approximately 13% relative to ISCC and 6% relative to PeNCache. In addition, content access delay decreased by approximately 3.5%.</p>