<p>The dissemination of fake news on social platforms represents a significant and pressing issue in contemporary society. Traditional approaches have employed various artificial intelligence (AI) technologies to verify news accuracy and provide explanations for the outcomes, achieving notable achievement in interpretable fake news detection (FND). This paper presents a fusion approach for FND on social media, addressing the challenge of limited interpretability, data fusion, and uncertainty in predictions. This method combines textual and visual information using pre-trained models EfficientNetB0 for image analysis and Electra and XLNet for text, to form a robust multimodal framework. Moreover, error level analysis (ELA) has been utilized to highlight the altered image aspects. This model employs ensemble soft voting to combine predictions from each model to make a final decision. To enhance interpretability, model-agnostic Shapley additive explanations (MASHAP) is utilized allowing the model to explain its predictions by identifying influential features within both text and images. Experiments conducted on three multimodal datasets MediaEval, CASIA, and Weibo demonstrate superior early detection capabilities, achieving accuracies of 95.9%, 95.7%, and 96.8%, respectively. This highlights the efficiency of the suggested method in detecting fake news, achieving a notably advanced level by integrating information across multiple modalities and employing a robust modality fusion technique.</p>