Abstract
The Social Internet of Things (SIoT) represents an evolution of the IoT where smart objects autonomously establish social-like relationships to facilitate collaboration. Ensuring security in such decentralized networks requires a robust trust evaluation mechanism. Traditional models often utilize static "weighted-sum" heuristics that fail to capture the dynamic and context-dependent nature of trust. This paper presents a time-aware trust evaluation model that utilizes a machine learning pipeline—combining k-means clustering for unsupervised labeling and Random Forest for classification. Evaluated on real-world traces from the CRAWDAD dataset, the proposed model achieves a classification accuracy of 98.8%, demonstrating its ability to effectively distinguish between trustworthy and malicious entities based on their social interaction history.
1. Introduction
As the Internet of Things (IoT) transitions into the Social Internet of Things (SIoT), smart objects—ranging from wearable devices to home hubs—increasingly interact through social-like paradigms such as friendships, shared ownership, and co-location. In these environments, objects must autonomously decide whether to collaborate with others. This decision-making process hinges on "trust," a metric that is inherently dynamic and influenced by past behaviors.
Existing trust evaluation frameworks typically rely on manually assigned weights for various parameters. These heuristics are difficult to tune and often fail to adapt to the evolving behaviors of nodes over time. There is a critical need for a model that can automatically determine the importance of social metrics and account for temporal changes in node behavior.
2. Proposed Methodology
The authors propose a three-stage pipeline designed to automate trust evaluation and adapt to the temporal dynamics of the SIoT network.
2.1 Feature Extraction
The model identifies four primary social trust metrics:
- Friendship Similarity (FS): Measures commonality in neighboring nodes.
- Community-of-Interest (CoI): Evaluates shared social or functional group memberships.
- Co-work Similarity (CWS): Quantifies the frequency of collaboration on common tasks.
- Cooperativeness (CoP): Uses an entropy-based function to detect interaction imbalances, which can signal malicious or inconsistent service.
2.2 Unsupervised Labeling and Classification
Given the lack of labeled "trustworthy" or "malicious" data in real-world SIoT traces, the model employs k-means clustering to group interactions into two classes. Subsequently, a Random Forest classifier is trained on these labels. This approach not only establishes a robust decision boundary but also provides "feature importance" scores, effectively replacing manual weighting with data-driven insights.
3. Experimental Evaluation
The model was validated using the CRAWDAD dataset (SIGCOMM 2009), which contains real-world traces of 76 nodes and over 18,000 interactions. The data includes friendship lists, interest groups, and message logs, which were processed into a feature matrix representing 5,776 interaction pairs.
4. Results and Analysis
The machine learning-driven approach yielded significant improvements over traditional models.
4.1 Classification Performance
The Random Forest model achieved an overall accuracy of 98.8%. Precision, recall, and F1-scores for both trustworthy and untrustworthy classes remained consistently between 0.97 and 1.0, indicating high reliability in identifying malicious actors.
4.2 Feature Importance and Temporal Dynamics
A key finding was the "death of manual weights." The Random Forest analysis revealed that Co-work Similarity (CWS) was the most influential factor, accounting for 60.75% of the trust decision. In contrast, Cooperativeness (CoP) contributed only 3.16% in this specific dataset.
Furthermore, the model demonstrated sensitivity to temporal changes. By tracking nodes over a 24-hour period, the authors observed how trust scores evolve. For instance, "Node 8" successfully built a positive reputation through consistent interactions, while "Node 19" lost trust over time due to deteriorating behavior. This highlights the model's ability to handle the "warm-up" period for new nodes and detect shifting patterns in established ones.
5. Conclusion
This paper demonstrates that a machine learning-driven, time-aware approach significantly enhances trust evaluation in SIoT. By moving away from rigid heuristics to a data-driven pipeline, the model achieves near-perfect accuracy while providing transparency into which social factors drive trust. Future work will focus on integrating environmental context (e.g., energy levels) and testing resilience against sophisticated "on-off" trust attacks.
Summary updated