COBRA: Context-aware Bernoulli Neural Networks for Reputation Assessment

Leonid Zeynalvand, Tie Luo, Jie Zhang

Abstract

Trust and reputation management (TRM) plays an increasingly important role in large-scale online environments such as multi-agent systems (MAS) and the Internet of Things (IoT). One main objective of TRM is to achieve accurate trust assessment of entities such as agents or IoT service providers. However, this encounters an accuracy-privacy dilemma as we identify in this paper, and we propose a framework called Context-aware Bernoulli Neural Network based Reputation Assessment (COBRA) to address this challenge. COBRA encapsulates agent interactions or transactions, which are prone to privacy leak, in machine learning models, and aggregates multiple such models using a Bernoulli neural network to predict a trust score for an agent. COBRA preserves agent privacy and retains interaction contexts via the machine learning models, and achieves more accurate trust prediction than a fully-connected neural network alternative. COBRA is also robust to security attacks by agents who inject fake machine learning models; notably, it is resistant to the 51-percent attack. The performance of COBRA is validated by our experiments using a real dataset, and by our simulations, where we also show that COBRA outperforms other state-of-the-art TRM systems.

Introduction

In the rapidly expanding worlds of the Internet of Things (IoT) and multi-agent systems, trust is the fundamental currency of interaction. Deciding whether a service provider will deliver on its promises requires robust reputation assessment. However, current Trust and Reputation Management (TRM) systems face a persistent hurdle: the accuracy-privacy dilemma.

Traditional systems typically fall into two categories. Systems with low privacy and high accuracy require agents to share full interaction histories (e.g., specific delivery times and locations), which exposes sensitive business or personal data. Conversely, systems with high privacy and low accuracy rely on simple numerical ratings (e.g., a 1-to-5 star rating). While ratings preserve privacy, they strip away vital context; a service provider might be highly reliable at noon but fail consistently at 5 PM. A single score hides these nuances, leading to poor trust predictions. Furthermore, most existing TRM systems are vulnerable to "51-percent attacks," where a majority of malicious agents collude to provide false testimonies.

The COBRA Framework

To resolve this dilemma, we propose COBRA (Context-aware Bernoulli Neural Network based Reputation Assessment). COBRA introduces two core innovations: model encapsulation and the use of Bernoulli Neural Networks (BNN).

Model Encapsulation

Instead of sharing raw interaction data or reductive ratings, each agent in the COBRA framework trains a local machine learning model—such as a Decision Tree, Naive Bayes, or Random Forest—on its own private experiences. When an agent seeks advice about a provider, the advising agents share these models rather than the data. This approach encapsulates context-aware knowledge (e.g., performance patterns based on time or location) without revealing the underlying raw data points, effectively bypassing the privacy leak.

Bernoulli Neural Networks (BNN)

The "brain" of the COBRA framework is the BNN, a specialized neural network designed to aggregate these shared models. Unlike standard artificial neural networks, the BNN is specifically optimized for inputs that are probabilities (the outputs of the encapsulated models).

The COBRA framework: agents share encapsulated models which are then aggregated by a Bernoulli Neural Network.

One of the most critical features of the BNN is its entropy-driven topology. The network uses information entropy to determine its internal connections, prioritizing inputs that are more "informative" (lower entropy). This allows the system to naturally filter out noise and malicious advice, as inconsistent or random testimonies from malicious agents result in higher entropy and are subsequently weighted less by the BNN.

Experimental Evaluation

We evaluated COBRA using both real-world data and simulated environments:

  • Real-World Dataset: We utilized a public dataset containing 30 million records of response times for 4,532 web services. This provided a realistic, sparse environment to test the accuracy of context-aware predictions.
  • Simulated Environment: To test security and robustness, we created a multi-agent simulation featuring a "majority-malice" scenario with 51 malicious agents and 49 legitimate ones.

Results and Discussion

COBRA was compared against several state-of-the-art benchmarks, including HABIT, TMSIoT, and the Beta Reputation System (BRS).

Superior Accuracy and Efficiency

As shown in our results, COBRA consistently achieved higher accuracy (ACC) and lower Root Mean Squared Error (RMSE) than its competitors. By preserving context through encapsulated models, COBRA-DT-B (Decision Tree with BNN) outperformed traditional rating-based systems significantly.

Performance metrics including accuracy comparison (a), prediction time (c), and convergence (e).

Beyond accuracy, the BNN architecture proved to be highly efficient. It demonstrated faster training times than standard "Dense" (fully-connected) neural networks and required less time for predictions as the number of agents in the system increased. The BNN also exhibited superior stability, with a smaller gap between training and validation accuracy, indicating high resistance to overfitting.

Resilience to 51-Percent Attacks

Perhaps most notably, COBRA proved remarkably robust against security threats. In simulations where malicious agents constituted 51% of the network, COBRA maintained an accuracy of over 85% in the vast majority of scenarios. This resilience is a direct result of the entropy-driven weighting, which allows the BNN to identify and ignore collusive, inconsistent advice even when it comes from the majority.

Conclusion

COBRA represents a significant step forward in Trust and Reputation Management by successfully navigating the accuracy-privacy dilemma. By combining model encapsulation with the probabilistic aggregation of Bernoulli Neural Networks, COBRA provides a secure, efficient, and context-aware framework for trust assessment. Future work will focus on standardizing context features across heterogeneous models and further optimizing the computational overhead for resource-constrained IoT devices.

Summary updated

Abstract

Trust and reputation management (TRM) plays an increasingly important role in large-scale online environments such as multi-agent systems (MAS) and the Internet of Things (IoT). One main objective of TRM is to achieve accurate trust assessment of entities such as agents or IoT service providers. However, this encounters an accuracy-privacy dilemma as we identify in this paper, and we propose a framework called Context-aware Bernoulli Neural Network based Reputation Assessment (COBRA) to address this challenge. COBRA encapsulates agent interactions or transactions, which are prone to privacy leak, in machine learning models, and aggregates multiple such models using a Bernoulli neural network to predict a trust score for an agent. COBRA preserves agent privacy and retains interaction contexts via the machine learning models, and achieves more accurate trust prediction than a fully-connected neural network alternative. COBRA is also robust to security attacks by agents who inject fake machine learning models; notably, it is resistant to the 51-percent attack. The performance of COBRA is validated by our experiments using a real dataset, and by our simulations, where we also show that COBRA outperforms other state-of-the-art TRM systems.
BibTeX
@inproceedings{zeynalvand2020cobra,
  author = {Leonid Zeynalvand and Tie Luo and Jie Zhang},
  title = {{COBRA: Context-aware Bernoulli Neural Networks for Reputation Assessment}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i05.6225},
}
Figure