COBRA: Context-Aware Bernoulli Neural Networks for Reputation Assessment

Leonid Zeynalvand, Tie Luo, Jie Zhang

Introduction

In decentralized ecosystems such as the Internet of Things (IoT) and multi-agent systems, establishing trust between peers is critical yet challenging. Traditional Trust and Reputation Management (TRM) systems often force a compromise between privacy and accuracy. Sharing detailed transaction logs provides high accuracy but compromises sensitive data, while simplified ratings preserve privacy at the cost of losing vital context. This "accuracy-privacy dilemma" is further complicated by the vulnerability of TRM systems to collusion attacks, where malicious agents provide deceptive feedback.

The COBRA (Context-Aware Bernoulli Neural Networks for Reputation Assessment) framework addresses these challenges by enabling agents to share local machine learning models instead of raw data. By aggregating these models through a specialized Bernoulli Neural Network (BNN), COBRA achieves state-of-the-art accuracy and robustness while maintaining data privacy.

The COBRA Framework

The COBRA architecture consists of two primary components: model encapsulation and BNN-based aggregation.

Model Encapsulation

Rather than transmitting raw interaction logs or scalar ratings, agents train local machine learning models (e.g., Decision Trees, Naive Bayes, or Random Forests) on their private experiences. When an agent seeks advice, the advisor shares the parameters of these local models. This approach encapsulates the nuances of past interactions—such as performance variations based on time or location—without revealing specific transaction details.

Bernoulli Neural Network (BNN)

The core of the COBRA framework is the Bernoulli Neural Network, designed specifically to process inputs that represent probabilities (Bernoulli random variables). Unlike standard dense neural networks, the BNN utilizes an entropy-based topology. It automatically weights inputs based on their information entropy, effectively prioritizing reliable advisors and discounting noisy or uncertain feedback. The network's 3-layer structure is dynamically configured based on the average entropy of the input data.

The COBRA architecture: Model encapsulation and BNN aggregation.

Experimental Evaluation

The researchers evaluated COBRA using a combination of real-world data and synthetic simulations:

  • Real-World Dataset: A public dataset comprising 30 million records from 4,532 web services and 142 users, focusing on Service Level Agreement (SLA) compliance.
  • Synthetic Simulations: A multi-agent environment with 100 agents, including a majority (51%) of malicious actors, to test resistance against collusion and "bad-mouthing" attacks.

Results and Analysis

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

Performance Metrics

COBRA demonstrated superior performance across all primary metrics. It achieved a peak accuracy (ACC) of 0.9117, outperforming HABIT (0.8952) and BRS (0.8403). Correspondingly, COBRA's Root Mean Squared Error (RMSE) was significantly lower at 0.2606, compared to 0.3178 for HABIT and 0.3512 for BRS.

Robustness and Efficiency

The BNN architecture proved to be more efficient than standard dense networks, showing faster training times and better resistance to overfitting. In adversarial simulations, COBRA maintained over 85% accuracy in 90% of scenarios, even under a 51-percent attack. This robustness stems from the BNN's ability to correlate honest first-hand evidence with the second-hand models provided by advisors.

Performance comparison of COBRA against baseline models across accuracy, error, and efficiency metrics.

Conclusion

COBRA represents a significant advancement in reputation management by leveraging model encapsulation and entropy-based neural networks. By treating models as a "middle ground" for knowledge transfer, it successfully navigates the trade-off between privacy and context-aware accuracy. Future research will focus on addressing the "cold start" problem in sparse networks and optimizing the computational overhead of local model training 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