We propose a novel context-based service framework to detect changes in social media images using image metadata. In this service-oriented architecture, each image is abstracted as a service. This method is unique in its exclusive reliance on metadata tags, which serve as non-functional attributes of the image service. We employ a transformer model trained on a large image metadata corpus to identify changes in image services. The training process is meticulously designed to focus solely on the metadata of image services falling within the relevant context, ensuring high precision. This targeted training ensures that the models are finely tuned to detect even subtle modifications pertinent to their respective categories, thereby improving the reliability and effectiveness of the framework. In this regard, a state-of-the-art transformer model, BERT is first trained on this corpus and then fine-tuned on a fact-checked, context-based dataset. Experiments are conducted on a subset of the Multimodal C4 dataset. Results demonstrate that the model’s effectiveness improves up to 20% when utilizing a context-based corpus, highlighting the value of targeted training in enhancing the accuracy of change detection.