This paper proposes a novel model called AKA-Fake for detecting fake news in a multimodal context. The model addresses the limitations of existing knowledge-enhanced methods by generating an adaptive knowledge subgraph through reinforcement learning. This subgraph captures relevant knowledge entities and their relationships, enabling more accurate detection of fake news. The model also incorporates a heterogeneous graph learning module to capture cross-modal correlations and a hierarchical modality-attentive graph pooling module to aggregate information from different modalities. The model is evaluated on three popular datasets and shows superior performance compared to existing baselines. The key contributions include the first investigation of adaptive knowledge learning for multimodal fake news detection, the proposal of the AKA-Fake model, and extensive experiments demonstrating its effectiveness. The model's ability to adaptively incorporate external knowledge and effectively model complex relationships between multimodal content and knowledge entities makes it a promising approach for fake news detection.This paper proposes a novel model called AKA-Fake for detecting fake news in a multimodal context. The model addresses the limitations of existing knowledge-enhanced methods by generating an adaptive knowledge subgraph through reinforcement learning. This subgraph captures relevant knowledge entities and their relationships, enabling more accurate detection of fake news. The model also incorporates a heterogeneous graph learning module to capture cross-modal correlations and a hierarchical modality-attentive graph pooling module to aggregate information from different modalities. The model is evaluated on three popular datasets and shows superior performance compared to existing baselines. The key contributions include the first investigation of adaptive knowledge learning for multimodal fake news detection, the proposal of the AKA-Fake model, and extensive experiments demonstrating its effectiveness. The model's ability to adaptively incorporate external knowledge and effectively model complex relationships between multimodal content and knowledge entities makes it a promising approach for fake news detection.