Automated Identification of Social Media Bots Using Deepfake Text Detection

Abstract

Social networks are playing an increasingly important role in modern society. Social media bots are also on the rise. Bots can propagate misinformation and spam, thereby influencing economy, politics, and healthcare. The progress in Natural Language Processing (NLP) techniques makes bots more deceptive and harder to detect. Easy availability of readily deployable bots empowers the attacker to perform malicious activities; this makes bot detection an important problem in social networks. Researchers have worked on the problem of bot detection. Most research focus on identifying bot accounts in social media; however, the meta-data needed for bot account detection is unavailable in many cases. Moreover, if the account is controlled by a cyborg (a bot-assisted human or human-assisted bot) such detection mechanisms will fail. Consequently, we focus on identifying bots on the basis of textual contents of posts they make in the social media, which we refer to as fake posts. NLP techniques based on Deep Learning appear to be the most promising approach for fake text detection. We employ an end-to-end neural network architecture for deep fake text detection on a real-world Twitter dataset containing deceptive Tweets. Our experiments achieve the state of the art performance and improve the classification accuracy by 2% compared to previously tested models. Moreover, our content-level approach can be used for fake posts detection in social media in real-time. Detecting fake texts before it gets propagated will help curb the spread of misinformation.

This work was supported in part by funds from NIST under award number 60NANB18D204, and from NSF under award number CNS 2027750, CNS 1822118 and from NIST, Statnett, Cyber Risk Research, AMI, ARL, and from DoE NEUP Program contract number DE-NE0008986.