<ul><li>Evidential Deep Learning is an efficient method for uncertainty quantification but is vulnerable to adversarially perturbed inputs, leading to overconfident errors.</li><li>Conflict-aware Evidential Deep Learning (C-EDL) is introduced as a post-hoc uncertainty quantification approach to enhance adversarial and out-of-distribution (OOD) robustness without retraining.</li><li>C-EDL generates diverse task-preserving transformations per input and calibrates uncertainty estimates by quantifying representational disagreement.</li><li>Experimental evaluation shows C-EDL outperforms existing EDL variants and baselines, with significant improvements in detecting OOD and adversarial inputs while maintaining high accuracy and low computational overhead.</li></ul>

Quantifying Adversarial Uncertainty in Evidential Deep Learning using Conflict Resolution

Discover more