Modeling and stabilizing perturbation propagation in multimodal learning via oscillatory state-space dynamics

Abstract

Multimodal learning systems are vulnerable to perturbations at various stages. Current robustness studies often overlook how disturbances spread across the pipeline. We propose reframing multimodal robustness as perturbation propagation and model it as a controlled state evolution system. To analyze fusion-path sensitivity, we introduce a Fusion-Path Adversarial Attack (FPAA). We propose a Dual-Oscillatory Mamba Fusion Network that integrates contraction-constrained state-space evolution with token-level and decision-level oscillatory regulation, limiting perturbation gain and preserving cross-modal coherence. Experiments show our framework outperforms external baselines in fusion-path robustness while maintaining competitive predictive performance. Ablation studies and stability measurements further support the contribution of each component.