Domain-Aware LoRA for Multimodal Fake News Detection
Abstract
Fake news detection is a critical artificial intelligence task. Early methods rely on unimodal information, while recent research favors multimodal and multi-domain learning for robust detection. Despite the progress, existing methods have three prominent limitations in multi-domain scenarios: insufficient domain-specific feature extraction, severe cross-domain data imbalance that biases models toward majority domains and impairs tail-domain performance, and static fusion strategies incapable of adapting to domain-dependent modality preferences, resulting in biased multimodal aggregation. This work proposes DALFND, a multi-domain multimodal fake news detection framework. It adopts domain-specific LoRA modules to extract discriminative domain knowledge and mitigate cross-domain feature bias for training-observed domains. Besides, a domain-aware dynamic fusion module is designed to capture domain-modality preferences and implement adaptive, stable cross-modal feature aggregation. Experiments on public benchmarks verify that DALFND achieves competitive and stable performance against mainstream baseline models in multi-domain fake news detection.