Deep Fake Reviews Spotting AI Synthesized Text in Dravidian Languages

The emergence of AI-written and fake reviews is a significant threat to online review websites. To counter this, we introduce a supervised learning-based solution with the use of Natural Language Processing (NLP) methods to identify human and AI-written reviews in Tamil and Malayalam. Our approach uses TF-IDF vectorization as the method for feature extraction and SMOTE for addressing class imbalance. We tried out TabPFN, LightGBM, and Isolation Forest and found that TabPFN performed the best with macro-average F1-scores of 74.79% (Tamil) and 67.83% (Malayalam). In contrast to BERT, which consumes a lot of labeled data and computational resources, TabPFN is adept at few-shot learning and thus is very effective for low-resource languages. LightGBM too performed well but needed to be carefully tuned for hyperparameters, whereas Isolation Forest was utilized for anomaly detection. Cross-validation validated TabPFN as the best-performing model with a mean F1-score of 71.31%, and it was a computationally efficient and reliable alternative to transformer-based models. Our study helps to boost trust in review sites by identifying AI-generated fake information effectively.