With the rapid development of multi-modal deepfake generation techniques, the malicious use and spread of forged audio-visual content pose severe security threats. However, most existing multi-modal deepfake detection methods rely on shallow feature concatenation or simple alignment, which fail to capture deep semantic correlations between modalities and thus lack robustness against unseen manipulations. To address these limitations, we propose MC-AVD, a generalizable audio-visual deepfake detection framework based on multi-modal collaboration. MC-AVD is built around three tightly coupled components. The modality-specific feature extraction module distills discriminative representations from audio and visual streams. These are refined by the cross-modal semantic alignment module, which maps heterogeneous features into a shared semantic space to enhance comparability and complementarity. Finally, the audio-visual semantic fusion module captures fine-grained semantic interactions across modalities, enabling the detection of subtle inconsistencies. On FakeAVCeleb, MC-AVD achieves up to 30.20% AUC improvement under cross-manipulation evaluation, demonstrating improved generalization to unseen forgery types.