A Multi-Faceted Approach to Trending Topic Attack Detection Using Semantic Similarity and Large-Scale Datasets

Twitter's widespread popularity has made it a prime target for malicious actors exploiting trending hashtags to disseminate harmful content. This study marks the first systematic exploration of semantic consistency in tweets to detect trending topic attacks. Unlike previous approaches, we emphasize the semantic aspect of tweets, leveraging advanced techniques such as semantic similarity estimation using WordNet and contextual understanding through Sentence-Transformers. To support this methodology, we curated large-scale, high-quality datasets comprising 7,000 Arabic and 28,000 English tweets, applying tailored preprocessing steps to ensure efficiency and accuracy. A novel data augmentation technique further enriched the quality and diversity of these datasets. We evaluated our approach using a comprehensive framework that assessed textual, image, and overall similarity. Five machine learning models - Random Forest, Decision Tree, K-Neighbors, Gradient Boosting, and XGBoost - were tested, with results benchmarked against nine baseline methods across different linguistic datasets and learning scenarios. Our approach demonstrated superior performance, achieving F1-scores of 96% for English and 97% for Arabic, with accuracy improvements ranging from 2% to 14% for English and 5% to 28% for Arabic. These results establish a new benchmark for detecting trending topic attacks across languages, highlighting the robustness and effectiveness of our method in combating malicious activities on social platforms.