Reliable Disentanglement Multi-view Learning Against View Adversarial Attacks

Xuyang Wang, Siyuan Duan, Qizhi Li, Guiduo Duan, Yuan Sun, Dezhong Peng

Introduction

Multi-view learning (MVL) has become a cornerstone of modern artificial intelligence, enabling systems to integrate information from diverse sources—such as combining LiDAR and camera data for autonomous navigation or aggregating various imaging modalities for medical diagnosis. To ensure these systems are safe for deployment, researchers have turned to Evidential Deep Learning (EDL). EDL allows models to quantify their own uncertainty, ideally signaling when data is too noisy or ambiguous to provide a reliable prediction.

However, a significant security gap remains. This paper identifies the Adversarial Unreliability Problem (AUP): current trusted MVL models are highly susceptible to adversarial attacks. These small, intentional perturbations not only cause the models to misclassify inputs but, more dangerously, often lead the models to become more confident in their incorrect predictions. This failure of uncertainty estimation poses a severe risk in safety-critical applications.

The RDML Framework

To address the AUP, the authors propose Reliable Disentanglement Multi-view Learning (RDML). RDML is a two-stage framework designed to move beyond passive fusion toward active defense by isolating and mitigating adversarial noise.

1. Perturbation-Insensitive Pretraining

The first stage involves training a robust "evidence extractor" using a combination of clean and adversarial data. This extractor serves as a "trusted guide" for the rest of the system, maintaining the ability to identify underlying evidence even when the input is compromised.

2. Evidential Disentanglement

Using the pretrained extractor as a supervisor, RDML employs a robustness-aware soft mask (implemented via Gumbel Softmax). This mask effectively disentangles the input features into two distinct streams: clean features and adversarial features. This separation is critical for preventing malicious perturbations from polluting the entire decision-making process.

3. Feature Recalibration and Attention

Unlike traditional defense mechanisms that simply discard noisy data, RDML attempts to maximize information utility:

  • Recalibration: An MLP-based module attempts to "fix" weak perturbations, recovering useful signals that would otherwise be lost.
  • Evidential Attention: For features that are too heavily compromised to be recalibrated, the model applies a view-level attention mechanism. This allows the system to dynamically ignore compromised views while focusing on the most reliable ones.

4. Dempster-Shafer Fusion

Finally, the processed information from all views is fused using Dempster-Shafer Theory. This mathematical framework allows the model to aggregate the "opinions" of different views into a single prediction that includes a valid, trustworthy uncertainty estimate.

Experimental Evaluation

The framework was rigorously tested across six diverse multi-view datasets, including PIE (facial images), Scene (indoor/outdoor categories), Leaves (plant species), NUS-WIDE (web images), MSRC-v5 (object recognition), and Fashion (apparel products).

Key Results

  • Superior Accuracy under Attack: When subjected to PGD (Projected Gradient Descent) attacks, RDML outperformed state-of-the-art methods by an average of 6.36%. On the PIE dataset specifically, RDML achieved 88.97% accuracy, significantly outperforming the next best evidence-based method (TMNR+AT), which reached only 71.32%.
  • Clean Data Performance: RDML does not sacrifice general performance for security. On clean datasets, it improved upon baseline models by an average of 1.41%, suggesting that its feature extraction process is inherently more robust.
  • Logical Uncertainty Estimation: A standout finding is that RDML’s uncertainty scores are physically meaningful. As the number of attacked views increases, the model’s uncertainty density shifts toward 1.0 (total uncertainty). In contrast, baseline models often remain highly confident in their errors, failing to signal the presence of an attack.

Discussion and Future Work

RDML demonstrates that disentanglement is a powerful tool for building trustworthy AI. By using a robust guide to separate noise from signal, the model can maintain both accuracy and honesty regarding its own limitations.

However, the authors note several areas for future exploration. The two-stage training process introduces additional computational complexity, and the current study focuses primarily on white-box attacks. Future research will need to address the scalability of this approach for high-resolution, real-time video streams and its effectiveness against a broader range of black-box adversarial strategies.

Summary updated

BibTeX
@inproceedings{wang2025reliable,
  author = {Xuyang Wang and Siyuan Duan and Qizhi Li and Guiduo Duan and Yuan Sun and Dezhong Peng},
  title = {{Reliable Disentanglement Multi-view Learning Against View Adversarial Attacks}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/720},
}
Figure