Adversarial attacks on a multimodal Alzheimer's disease detection system reveal complex interdependences between...

Area of Science:

• Artificial Intelligence
• Neuroscience
• Medical Imaging

Background:

• Multi-modal models integrating neuroimaging and clinical data are state-of-the-art for Alzheimer's disease detection.
• The adversarial robustness of these complex models is not well understood.

Purpose of the Study:

• To systematically investigate adversarial vulnerability in a multi-modal Alzheimer's detection model (CogniNetMM).
• To analyze vulnerability across different data fusion configurations (MRI only, clinical only, joint).
• To introduce and utilize a mean attack framework to probe decision boundary properties.

Main Methods:

• Employed Fast Gradient Sign Method and DeepFool adversarial attack algorithms.
• Evaluated three data configurations: MRI, clinical variables, and joint fusion.
• Introduced a mean attack framework to analyze fixed-direction perturbations.

Main Results:

• Identified 'modality collapse' during training, where the fusion layer suppresses MRI data, reducible by larger batch sizes and stratified sampling.
• Joint multi-modal attacks showed higher success rates than unimodal attacks, particularly with DeepFool near the decision boundary.
• The mean attack confirmed that increased vulnerability is a structural property of the fused decision boundary.

Conclusions:

• Heterogeneous data fusion in multi-modal models introduces emergent adversarial vulnerabilities not predicted by unimodal analysis.
• Standard training on imbalanced medical data may lead to silent modality suppression, impacting model reliability.
• Understanding and mitigating these emergent vulnerabilities is crucial for robust automated disease detection.