A Review of Deepfake Tweet Detection Using Sentiment Majority Voting Classifier with BERT and Random Forest

Abstract

Deepfake content on social media, particularly tweets, poses significant challenges to information authenticity, often manipulating public opinion and spreading misinformation. This paper reviews a novel approach to detecting deepfake tweets by integrating a sentiment majority voting classifier with transfer learning-based feature engineering. Leveraging BERT for contextual embeddings and Random Forest for robust classification, the proposed system enhances detection accuracy and interpretability. The methodology utilizes a dataset of labeled tweets (human vs. bot) from Kaggle, achieving superior performance over traditional methods like Decision Trees, SVM, KNN, Logistic Regression, and LSTM, with a Turnitin similarity score of 18%. Key findings include improved accuracy, precision, recall, and F1-score, alongside insights into feature importance for transparency. This review highlights the system’s potential to combat misinformation on social media, its implications for stakeholders, and limitations such as its focus on English tweets. Future directions include expanding to multilingual datasets and deeper XAI exploration. This approach contributes to reliable sentiment analysis, fostering trust in digital communication platforms.