With the rapid development of social media and news platforms, the spread of fake news has had an increasingly significant impact on society. Existing approaches for fake news detection that address domain shift predominantly rely on mining the relationship between image and textual features. However, these methods neglect the importance of frequency features in cross-domain generalization detection. Effectively extracting multi-scale and directional frequency features, and properly disentangling domain-invariant features of news and forgery-type-invariant features from Image-Frequency (I-F) and Image-Text (I-T) features, remain significant challenges in addressing domain shift. In this paper, we propose a Frequency-Aware Robust Multimodal Fake News detection framework (FAR-MFN), which detects fake news by learning domain-invariant features of news and forgery-type-invariant features. Specifically, FAR-MFN disentangles I-F and I-T features using Enhanced MMoE (EMMoE), thereby extracting invariant features. Additionally, to effectively capture multi-scale and directional frequency features, we introduce a frequency encoder based on Frequency-Aware Window Attention (FAWA), which integrates both intra-band and inter-band self-attention mechanisms to comprehensively capture spectral features. Extensive experiments demonstrate that FAR-MFN achieves state-of-the-art performance on the cross-domain fake news detection benchmark.