A Fake News Detection Technology Based on Adversarial Contrast Under Few-Shot Samples

The development and widespread adoption of the Internet have made online social platforms a bridge for communication between users, with hundreds of millions expressing opinions and exchanging views daily. However, this is accompanied by the spread of a vast amount of unverified misinformation. A particular challenge is the emergence of cross-lingual breaking misinformation, where the same textual content in different languages is published on social platforms across various countries, which further complicates feature extraction. To address this, this paper proposes an adversarial contrastive learning method for fake news detection. It maps cross-lingual texts into a shared semantic space and employs a capsule network to generalize class-specific features. Utilizing a contrastive learning framework, the method calculates the distance between target and source data. Finally, noise is incorporated to further enhance the robustness of fake news detection in few-shot scenarios. Experiments on Twitter and Weibo datasets show that the proposed method improves accuracy by 4.8% and 3.0%, respectively, compared to the best baseline method.