Since the deployment of the Internet of Things (IoT), it has transformed everyday life by enabling intelligent environments that improve efficiency and automate services in domains such as agriculture, healthcare, smart cities, and industry. However, the rapid proliferation of IoT devices has introduced significant security challenges, largely driven by the heterogeneity of devices, resource constraints, and the increasing exposure of network communications. This work proposes a lightweight security protection mechanism for IoT networks based on trust modelling. The proposed approach integrates machine learning techniques to evaluate IoT node behavior using network-layer (Layer 3) traffic features under different labeling granularities, including binary, categorical, and subcategorical classifications. By focusing on network-layer observations, the model remains applicable across heterogeneous IoT devices while preserving a low computational footprint. In addition, the Common Vulnerability Scoring System (CVSS) is incorporated as a standardized vulnerability severity metric, enabling the integration of probabilistic security evidence with contextual information about potential impact. This combination allows the estimation of trust to reflect not only the likelihood of anomalous behavior but also its associated severity. Experimental evaluation was conducted using a representative IoT traffic dataset, multiple preprocessing strategies, and several classical machine learning models. The results demonstrate that aggregating traffic-based intrusion detection outputs with vulnerability severity metrics enables a more robust, flexible, and interpretable trust estimation process. This approach supports the early identification of potentially compromised nodes while maintaining scalability and efficiency, making it suitable for deployment in heterogeneous IoT environments.