PaTAS: A Neural Network Trust Propagation Framework Based on Subjective Logic

The PaTAS framework propagates trust in parallel within neural networks based on subjective logic, giving each prediction a trust opinion comprising belief, disbelief and uncertainty. On real and adversarial datasets, it can flag accurate predictions that rely on parameters learned from contaminated data, detect adversarial patch inputs without trigger information, and audit training label contamination when no test data is available. Compared with baselines such as output uncertainty, trust propagation supplements model-side signals, and the mechanism can be extended to convolutional networks.