PaTAS: A Framework for Trust Propagation in Neural Networks Using Subjective Logic

Researchers introduce PaTAS (Parallel Trust Assessment System), a framework that uses Subjective Logic to measure and propagate trust through neural networks alongside standard computation. The system identifies reliability gaps and adversarial vulnerabilities that traditional metrics like accuracy fail to detect, offering a foundation for deploying AI safely in critical applications.

PaTAS addresses a fundamental gap in AI deployment: conventional performance metrics do not capture whether a model's confidence aligns with actual reliability. By introducing Trust Nodes and Trust Functions that operate parallel to standard neural computation, the framework enables quantifiable assessment of trustworthiness across the AI lifecycle. This distinction matters significantly because high accuracy does not guarantee robustness under adversarial attacks, data poisoning, or distribution shifts—scenarios increasingly relevant as AI systems enter high-stakes domains like healthcare, autonomous vehicles, and financial systems.

The research builds on growing recognition that interpretability and reliability verification are prerequisites for enterprise and regulatory adoption of neural networks. Traditional approaches evaluate models in isolation; PaTAS extends trust reasoning to instance-specific predictions, revealing when model confidence diverges from actual accuracy. The Parameter Trust Update mechanism during training and Inference-Path Trust Assessment at deployment provide continuous monitoring capabilities. This aligns with regulatory trends toward explainable AI and risk quantification, particularly as governments establish frameworks requiring transparency in automated decision-making.

For stakeholders developing mission-critical AI systems, PaTAS offers a scalable approach to identify failure modes without extensive manual auditing. Organizations deploying neural networks in regulated industries can use the framework to document reliability across varied conditions, reducing liability and building stakeholder confidence. The ability to distinguish benign from adversarial inputs has direct applications in fraud detection, cybersecurity, and safety-critical systems. As AI adoption accelerates, frameworks that quantify trust rather than merely report accuracy will become competitive differentiators in enterprise deployments.

• →PaTAS uses Subjective Logic to propagate and measure trust through neural networks independently of accuracy metrics.
• →The framework identifies reliability gaps and adversarial vulnerabilities that conventional metrics fail to detect.
• →Parameter Trust Updates during training and Inference-Path Trust Assessment enable continuous monitoring of model reliability.
• →Experimental results demonstrate PaTAS distinguishes between benign and adversarial inputs with interpretable, symmetric trust estimates.
• →The framework addresses regulatory and enterprise demand for transparent, quantifiable AI reliability in safety-critical applications.