Contrastive Domain Adaptation for Authenticity Detection in Chat Record Screenshots

Malicious actors use realistic social software interface simulators to forge deceptive screenshots, spread rumors, and manipulate public opinion, which seriously disrupts social order. At present, the verification of the authenticity of screenshots mainly relies on specialists' review of complex details such as icon size and text spacing. If the screenshots come from different devices, it will exponentially increase the difficulty of expert identification. In response to the issue of discerning the authenticity of chat record screenshots across various devices, we propose an automated framework and detection model, which aims to maintain the accuracy and efficiency in discerning the authenticity of chat record screenshots from different devices. We collect sample data from different devices by generating various text conversations through ChatGPT. We designed a solution that combines cross-domain learning and contrastive learning concepts. By fine-tuning with two loss functions, the model has improved its performance in identifying same-category samples across different device domains and different-category samples within the same device domain. The model has a detection accuracy of 98% for chat record screenshots with high reliability.