Conflict-aware Evidential Deep Learning

• Conflict-aware Evidential Deep Learning (C-EDL) is a framework that explicitly quantifies and mitigates evidence conflict to improve uncertainty estimation in deep models.
• It integrates methods like post-hoc adjustments and architecture-level DSCR to stabilize predictions in multi-view, incomplete, and adversarial scenarios.
• Empirical studies show that C-EDL achieves state-of-the-art performance in robust classification and detection, reducing overconfident mispredictions in challenging settings.

Conflict-aware Evidential Deep Learning (C-EDL) encompasses a family of methods designed to improve the robustness and uncertainty quantification of Evidential Deep Learning (EDL) models, especially under data regimes characterized by conflicting, incomplete, adversarial, or out-of-distribution (OOD) evidence. These approaches explicitly measure representational disagreement—either among input transformations or multi-view observations—and calibrate uncertainty scores and predictions accordingly. C-EDL instantiations include lightweight, post-hoc adjustments for pre-trained EDL classifiers as well as integrated, architecture-level conflict resolution within multi-view settings. They achieve state-of-the-art performance in OOD and adversarial detection, and in robust multi-view classification under missingness and view corruption (Chen et al., 2024, Barker et al., 6 Jun 2025).

1. Evidential Deep Learning: Foundations and Limitations

EDL provides a non-Bayesian framework for uncertainty estimation by modeling class predictions as parameters of a Dirichlet distribution. Given input $x \in \mathbb{R}^d$ and $K$ classes:

• The EDL network predicts non-negative evidence $e_k(x) \geq 0$. Dirichlet parameters $\alpha_k = e_k(x) + 1$.
• The total evidence $S(x) = \sum_{k=1}^K \alpha_k$.
• The belief mass $b_k(x) = \frac{e_k(x)}{S(x)}$ and the remaining uncertainty $u(x) = \tfrac{K}{S(x)}$.
• The expected predictive probability $E[p_k] = \alpha_k / S(x)$.

Training minimizes an “evidential loss” that combines a data fidelity term and a KL-divergence regularizer toward the Dirichlet uniform prior, penalizing overconfident mispredictions.

However, standard EDL is vulnerable to overconfident misclassification in OOD and adversarial settings. A single forward pass yields high evidence even for invalid or corrupted inputs, and traditional Dempster–Shafer (DS) fusion amplifies this problem in multi-view or transformation-ensemble scenarios due to its sensitivity to conflict (Barker et al., 6 Jun 2025, Chen et al., 2024).

2. Conflict in Evidential Fusion: Formalization and Impact

In DS theory, each evidence source provides a basic belief assignment (BBA) $m: 2^\Theta \rightarrow [0,1]$ over the frame of discernment $\Theta$. The DS combination rule fuses two BBAs $K$0 and $K$1 as:

• The conflict mass $K$2.
• The fused mass for $K$3, $K$4.

When sources are highly contradictory ($K$5), the denominator $K$6 approaches zero, causing numerical instability, unpredictable magnification of uncertainty, and degraded uncertainty quality. In incomplete multi-view classification, imputation errors can induce frequent moderate-to-severe conflicts during evidence fusion, undermining confidence calibration and model reliability (Chen et al., 2024).

This motivates conflict-aware mechanisms that explicitly measure and respond to evidence disagreement.

3. Conflict-aware Dempster–Shafer Combination Rule (DSCR) in Multi-View Learning

The Alternating Progressive Learning Network (APLN) and its DSCR represent a principled integration of conflict-aware EDL in multi-view, incomplete-data regimes (Chen et al., 2024). The DSCR operates as follows:

• For singleton opinions $K$7 and $K$8, conflict is $K$9.
• The unnormalized fused belief and uncertainty are:
  • $e_k(x) \geq 0$0
  • $e_k(x) \geq 0$1
• Both $e_k(x) \geq 0$2 and $e_k(x) \geq 0$3 are down-weighed by $e_k(x) \geq 0$4: $e_k(x) \geq 0$5, $e_k(x) \geq 0$6
• Final normalization: $e_k(x) \geq 0$7, $e_k(x) \geq 0$8, $e_k(x) \geq 0$9

This formulation ensures that strong conflict (large $\alpha_k = e_k(x) + 1$0) shrinks fused belief evidence and transfers mass to uncertainty, preventing instability and yielding a more reliable combined opinion under incompleteness and conflict.

Within APLN, multi-view data proceeds through three learning phases: coarse imputation and latent alignment (UMAE-F), progressive evidence learning (UMAE-V), and joint end-to-end optimization (UMAE-J), always fusing evidence via DSCR to stabilize both training and inference.

4. Post-hoc Conflict-Aware EDL (C-EDL) for OOD and Adversarial Detection

A distinct instantiation of C-EDL provides a post-hoc, lightweight approach for uncertainty calibration in standard EDL classifiers without retraining (Barker et al., 6 Jun 2025). The method operates as follows:

• For each test sample $\alpha_k = e_k(x) + 1$1, generate $\alpha_k = e_k(x) + 1$2 task-preserving, metamorphic transformations $\alpha_k = e_k(x) + 1$3.
• For each transformed input, obtain Dirichlet parameters $\alpha_k = e_k(x) + 1$4 from the pre-trained EDL network.
• Compute conflict over the evidence set $\alpha_k = e_k(x) + 1$5 using:
  • Intra-class variability: $\alpha_k = e_k(x) + 1$6
  • Inter-class contradiction: $\alpha_k = e_k(x) + 1$7 (see source for full detail)
  • Total conflict $\alpha_k = e_k(x) + 1$8, with $\alpha_k = e_k(x) + 1$9
• Average and decay evidence via $S(x) = \sum_{k=1}^K \alpha_k$0, $S(x) = \sum_{k=1}^K \alpha_k$1.
• Compute final uncertainty $S(x) = \sum_{k=1}^K \alpha_k$2 and use it for abstention or OOD/adversarial flagging: predict if $S(x) = \sum_{k=1}^K \alpha_k$3; otherwise abstain.

This framework uses conflict as a trigger to reduce posterior confidence, selectively elevating uncertainty where transformations strongly disagree, and suppressing overconfident errors on anomalous inputs.

5. Optimization and Loss Formulations

In the multi-view APLN setting with DSCR (Chen et al., 2024), the objective comprises:

• EDL evidence loss: $S(x) = \sum_{k=1}^K \alpha_k$4 (with digamma $S(x) = \sum_{k=1}^K \alpha_k$5)
• KL divergence to a uniform Dirichlet: $S(x) = \sum_{k=1}^K \alpha_k$6
• Conflict consistency loss: for every view pair $S(x) = \sum_{k=1}^K \alpha_k$7, compute a divergence $S(x) = \sum_{k=1}^K \alpha_k$8, and optimize the mean conflict degree $S(x) = \sum_{k=1}^K \alpha_k$9, then $b_k(x) = \frac{e_k(x)}{S(x)}$0
• ELBO regularization for latent imputation using a VAE

Sampling for missing views in APLN is stochastic: the VAE samples $b_k(x) = \frac{e_k(x)}{S(x)}$1 latent codes for each missing view, evidence is averaged before forming Dirichlet parameters, thus propagating uncertainty from missingness explicitly into the fused predictive distribution.

For post-hoc C-EDL (Barker et al., 6 Jun 2025), only inference phase computation is required, and no training loss modification is imposed.

6. Empirical Validation and Comparative Performance

C-EDL methods achieve consistent state-of-the-art performance across both incomplete multi-view settings and OOD/adversarial detection tasks.

In incomplete-view multi-view classification (Chen et al., 2024):

• Datasets: YaleB, Handwritten, ROSMAP, BRCA, Scene15, NUS-Wide.
• When missingness rate increases ($b_k(x) = \frac{e_k(x)}{S(x)}$2 up to 0.5), APLN+DSCR achieves highest accuracy (e.g., Handwritten at $b_k(x) = \frac{e_k(x)}{S(x)}$3: 97.05% vs UIMC’s 97.00%; ROSMAP at $b_k(x) = \frac{e_k(x)}{S(x)}$4: 72.97% vs 71.43%).
• On conflict test splits (e.g., 40% of samples with cross-class view swaps), APLN+DSCR maintains accuracy within 1–2% of non-conflict performance, while standard DS fusion suffers accuracy drops up to 5%.
• Uncertainty metrics improve (average $b_k(x) = \frac{e_k(x)}{S(x)}$5 decreases with more coherent evidence), and accuracy variance is reduced.

In OOD and adversarial detection (Barker et al., 6 Jun 2025):

• Across MNIST and CIFAR-10 tasks, C-EDL retains >94% ID coverage, reduces OOD coverage by up to 55%, and adversarial coverage by up to 90% compared to standard EDL.
• For example, under severe attack (MNIST$b_k(x) = \frac{e_k(x)}{S(x)}$6FashionMNIST, L2-PGD, $b_k(x) = \frac{e_k(x)}{S(x)}$7): EDL retains 52.21% of adversarial samples, C-EDL only 15.51%; ID coverage remains high (EDL 96.61%, C-EDL 94.18%).
• Runtime overhead is minimal ($b_k(x) = \frac{e_k(x)}{S(x)}$8 EDL), as only T ($b_k(x) = \frac{e_k(x)}{S(x)}$9) forward passes and lightweight evidence statistics are required.

Method	ID Coverage	OOD Coverage	Adv Coverage (L2-PGD, $u(x) = \tfrac{K}{S(x)}$0)
Standard EDL	96.61%	2.52%	52.21%
C-EDL	94.18%	1.77%	15.51%

These results demonstrate that C-EDL mechanisms, whether integrated (DSCR) or post-hoc, robustly prevent overconfident false predictions in high-conflict, high-uncertainty, and adversarial contexts without sacrificing in-distribution performance.

7. Significance and Theoretical Implications

Conflict-aware Evidential Deep Learning establishes a general methodology for enhancing epistemic uncertainty quantification in neural models:

• By explicitly quantifying and attenuating conflict, it stabilizes evidence aggregation in both multi-view and transformation-based ensembles.
• It is agnostic to architecture and can be used either as a training-integrated module (as in DSCR/APLN) or inference-only post-processing (as in C-EDL).
• The theoretical formulations remain consistent with Dempster–Shafer subjective logic, and provide analytic conflict metrics with monotonicity guarantees.
• The negligible computational overhead and the empirical state-of-the-art improvement in both robustness and uncertainty calibration distinguish C-EDL as a general-purpose uncertainty amplification strategy.

A plausible implication is that conflict quantification—via statistical divergence measures or DS-style combiners—could become standard for uncertainty adjustment in other evidential and Bayesian deep learning domains, especially as deployment in real-world safety-critical applications increases (Chen et al., 2024, Barker et al., 6 Jun 2025).