Introduction
The evolution of the Internet of Things (IoT) has led to the emergence of the Social Internet of Things (SIoT). In this paradigm, devices are not merely passive sensors but autonomous agents capable of establishing social-like relationships. These "social" objects can request services, share data, and collaborate based on their network ties. However, this autonomy introduces significant security risks. In a decentralized, massive-scale network, identifying whether a peer is trustworthy or malicious is a critical challenge.
Existing trust models often rely on simple linear equations that struggle to capture the non-linear, complex dynamics of social interactions. They frequently overlook the nuances of social context—such as shared ownership, location, or manufacturer—and fail to adequately account for the temporal decay of trust.
The Trust-SIoT Framework
To address these challenges, researchers developed Trust-SIoT, a multi-layered framework that treats trust evaluation as a classification problem solved by an Artificial Neural Network (ANN). Unlike traditional heuristic models, Trust-SIoT leverages five distinct metrics to assess a trustee's credibility:
- Direct Trust Metric (DTM): Calculated using Bayesian inference, this metric weighs successful versus unsuccessful interactions. It incorporates a decay factor to ensure recent behavior carries more weight than historical data.
- Reliability (R): This measures the "fairness" of an object's recommendations about other peers.
- Benevolence (B): This reflects the network's collective belief in an object's recommendations.
- Recommendations (RTM): This aggregates feedback from credible neighbors (those with high reliability and benevolence) to form a second-hand opinion.
- Context-aware Degree of Relationship (C-DoR): A key innovation of this work, C-DoR uses Knowledge Graph Embedding (KGE)—specifically the RotL model—to map complex social links (e.g., "owned by the same user") into mathematical vectors. This allows the system to quantify "social similarity" between devices.
These metrics are fed into an ANN that learns optimal weights, ultimately classifying devices as Trustworthy, Neutral, or Untrustworthy.
Experimental Setup and Datasets
Evaluating trust in SIoT requires datasets that combine social relationships with trust ratings. The authors synthesized a realistic environment by "grafting" social relationship types from the Social-IoT Dataset (based on real-world data from Santander, Spain) onto trust-labeled data from:
- Advogato: An online social network where users certify trustworthiness.
- Bitcoin-Alpha: A "web-of-trust" network used to prevent fraud in cryptocurrency exchanges.
Key Results and Insights
The Trust-SIoT framework was benchmarked against Guardian, a state-of-the-art trust heuristic. The results demonstrated significant improvements across several dimensions:
- Improved Accuracy: On the Advogato dataset, Trust-SIoT achieved an F1-score of 0.859, outperforming the baseline by 3.6%.
- Reduced Error: The model reduced Mean Absolute Error (MAE) by 18.7% and Mean Squared Error (MSE) by 20.2% compared to the baseline.
- Robustness: Trust-SIoT maintained stable performance even when training data was reduced to 40% of the total, whereas the baseline model's accuracy degraded significantly.
A primary insight is that trust in SIoT is inherently non-linear. By using an ANN, the model can identify subtle behavioral patterns that simple arithmetic formulas miss. Furthermore, the inclusion of social context via KGE proved that a device's "social circle"—its manufacturer or owner—is a powerful predictor of its likely behavior.
Conclusion
Trust-SIoT represents a significant step toward securing autonomous social networks of devices. By moving beyond linear heuristics and incorporating deep learning with knowledge graph embeddings, the framework provides a more resilient and accurate method for identifying malicious actors. Future research will focus on testing the model against sophisticated attacks like Sybil or "on-off" behaviors and evaluating its performance in highly dynamic environments where network topology changes rapidly.
Summary updated