Cryptify: Identifying Crypto Disinformation in Tweets Using Computational Techniques

Abstract

Disinformation in cryptocurrencies is a major concern in social media nowadays, especially in the platform X which is one of the most popular social media platforms used by crypto enthusiasts. Being a trending topic for the past decade, this has given rise to various forms of disinformation and manipulation which significantly impact both investors and the global cryptocurrency market. Detecting disinformation spread in cryptocurrency tweets is crucial due the fraudulent activities conducted such as disinformation campaigns, pump and dump scams, market manipulation. Addressing this prevailing issue is vital to protect investors, for maintaining market integrity and for ensuring transparency throughout the cryptocurrency space. This research project involved developing a computational approach to automatically detect disinformation in cryptocurrency tweets by classifying tweet data passed into model to be either ‘disinformed’ or ‘legitimate’ based on the context of the data provided into the model. The proposed system, Cryptify was implemented by fine-tuning a GPT 3.5 Turbo model with the goal of fine-tuning the base model with its low accuracy edge cases improving the model and making sure to build a specialized model that’ll be far more accurate than the base model. Benchmarks were also conducted on the base model, the base model with prompting and the fine-tuned model, where the base model with prompting showed an accuracy of 90% compared to 85% of the base model and the fine-tuned model outperformed every other model with an accuracy of 92% along with other metrics such as Precision, Recall and F1 Score all outperforming the metrics of the fine-tuned GPT 3.5 Turbo model.