Introduction
Multi-view learning (MVL) has become a cornerstone for tasks requiring the integration of information from diverse sources, such as multi-modal sensors in autonomous vehicles or various medical imaging modalities. Recent advancements in "trusted" MVL leverage evidential learning to provide not just predictions, but also uncertainty estimates, allowing models to signal when data is conflicting or unreliable.
However, a critical vulnerability known as the Adversarial Unreliability Problem (AUP) threatens these systems. Standard trusted models often assume data integrity. When an attacker introduces adversarial perturbations—such as those generated by Projected Gradient Descent (PGD)—into even a single view, model performance collapses. More alarmingly, the model's uncertainty mechanism fails; it often becomes overconfident in its erroneous predictions, providing low uncertainty scores for incorrect results. This behavior is unacceptable for safety-critical applications where an "I don't know" response is safer than a confident error.
The RDML Framework
To address the AUP, we propose Reliable Disentanglement Multi-view Learning (RDML). This framework is designed to restore trust by explicitly separating clean information from malicious noise.
The RDML architecture consists of several key components:
- Perturbation-Insensitive Pretraining: An evidence extractor is pretrained on a mixture of clean and adversarial samples. This ensures the base component can provide stable category-level evidence even when the input is under attack.
- Evidential Disentanglement Learning: Using the pretrained extractor as a guide, the model generates a "robustness-aware soft mask." This mask acts as a filter, separating each view into a clean feature part and an adversarial feature part, preventing malicious signals from propagating into the final decision-making layers.
- Feature Recalibration: Instead of simply discarding adversarial components, RDML uses a dedicated module to "repair" weak perturbations. This process turns potentially harmful noise back into useful, discriminative features, maximizing the information extracted from every view.
- View-level Evidential Attention: For "stubborn" adversarial noise that cannot be repaired, the model applies an attention mechanism. This tells the fusion layer to ignore or down-weight the compromised dimensions of a specific view.
- Dempster Rule-based Fusion: Finally, the "opinions" from all views are combined using Dempster-Shafer theory. This approach treats each view's output as a piece of evidence with its own mass of belief and uncertainty, producing a final classification and a reliable uncertainty score.
Experimental Evaluation
RDML was evaluated across six diverse multi-view datasets: PIE and MSRC (image recognition), Scene and Fashion (categorization), Leaves (botany), and NUS-WIDE (large-scale web images). We simulated real-world sensor tampering using "random view attacks," where adversarial noise is injected into a subset of available views.
Key Results
RDML demonstrates superior robustness compared to state-of-the-art (SOTA) methods, outperforming the second-best models by an average of 6.36% in adversarial settings. On the PIE dataset, while previous models like ECML saw a 52% drop in accuracy under attack, RDML maintained 88.97% accuracy—a mere 9% decrease from its clean-data performance.
Beyond accuracy, RDML restores the reliability of uncertainty estimates. As shown in the uncertainty density plots, RDML's uncertainty scores correctly increase when views are compromised, providing a vital "fail-safe" signal. In contrast, baseline models often show decreased uncertainty under attack, making them dangerously overconfident in their errors. Furthermore, RDML improves performance on clean data by an average of 1.41% over baselines, suggesting that the disentanglement process enhances general feature extraction.
Conclusion and Future Work
RDML effectively mitigates the Adversarial Unreliability Problem by disentangling and recalibrating multi-view features. While highly effective against PGD attacks, future work will focus on addressing "stubborn" features that resist repair and exploring the framework's resilience against a broader range of attacks, including black-box and physical adversarial patches. Additionally, optimizing the computational overhead of the multi-stage training process remains a priority for real-time deployment.
Summary updated