PaTAS: Trust Propagation in Neural Networks via Subjective Logic
The Parallel Trust Assessment System (PaTAS) introduces a novel framework that employs Subjective Logic to assess and disseminate trust within neural networks. It functions in conjunction with conventional computation by utilizing Trust Nodes and Trust Functions, managing trust related to inputs, parameters, and activations. A mechanism for Parameter Trust Update enhances reliability throughout the training process, while Inference-Path Trust Assessment evaluates trust specific to instances during inference. Tests conducted on both real-world and adversarial datasets demonstrate its efficacy.
Key facts
- PaTAS stands for Parallel Trust Assessment System.
- It uses Subjective Logic (SL) for trust modeling.
- Trust Nodes and Trust Functions propagate trust across the network.
- Parameter Trust Update mechanism refines reliability during training.
- Inference-Path Trust Assessment (IPTA) computes instance-specific trust.
- Experiments were conducted on real-world and adversarial datasets.
- The framework addresses trustworthiness in safety-critical AI applications.
- Conventional metrics like accuracy fail to capture uncertainty.
Entities
—