Dataset Information

TweepFake: About detecting deepfake tweets.

ABSTRACT: The recent advances in language modeling significantly improved the generative capabilities of deep neural models: in 2019 OpenAI released GPT-2, a pre-trained language model that can autonomously generate coherent, non-trivial and human-like text samples. Since then, ever more powerful text generative models have been developed. Adversaries can exploit these tremendous generative capabilities to enhance social bots that will have the ability to write plausible deepfake messages, hoping to contaminate public debate. To prevent this, it is crucial to develop deepfake social media messages detection systems. However, to the best of our knowledge no one has ever addressed the detection of machine-generated texts on social networks like Twitter or Facebook. With the aim of helping the research in this detection field, we collected the first dataset of real deepfake tweets, TweepFake. It is real in the sense that each deepfake tweet was actually posted on Twitter. We collected tweets from a total of 23 bots, imitating 17 human accounts. The bots are based on various generation techniques, i.e., Markov Chains, RNN, RNN+Markov, LSTM, GPT-2. We also randomly selected tweets from the humans imitated by the bots to have an overall balanced dataset of 25,572 tweets (half human and half bots generated). The dataset is publicly available on Kaggle. Lastly, we evaluated 13 deepfake text detection methods (based on various state-of-the-art approaches) to both demonstrate the challenges that Tweepfake poses and create a solid baseline of detection techniques. We hope that TweepFake can offer the opportunity to tackle the deepfake detection on social media messages as well.

SUBMITTER: Fagni T

PROVIDER: S-EPMC8118345 | biostudies-literature |

REPOSITORIES: biostudies-literature

ACCESS DATA

Similar Datasets

Project description:BACKGROUND:Mental disorders have become a major concern in public health, and they are one of the main causes of the overall disease burden worldwide. Social media platforms allow us to observe the activities, thoughts, and feelings of people's daily lives, including those of patients suffering from mental disorders. There are studies that have analyzed the influence of mental disorders, including depression, in the behavior of social media users, but they have been usually focused on messages written in English. OBJECTIVE:The study aimed to identify the linguistic features of tweets in Spanish and the behavioral patterns of Twitter users who generate them, which could suggest signs of depression. METHODS:This study was developed in 2 steps. In the first step, the selection of users and the compilation of tweets were performed. A total of 3 datasets of tweets were created, a depressive users dataset (made up of the timeline of 90 users who explicitly mentioned that they suffer from depression), a depressive tweets dataset (a manual selection of tweets from the previous users, which included expressions indicative of depression), and a control dataset (made up of the timeline of 450 randomly selected users). In the second step, the comparison and analysis of the 3 datasets of tweets were carried out. RESULTS:In comparison with the control dataset, the depressive users are less active in posting tweets, doing it more frequently between 23:00 and 6:00 (P<.001). The percentage of nouns used by the control dataset almost doubles that of the depressive users (P<.001). By contrast, the use of verbs is more common in the depressive users dataset (P<.001). The first-person singular pronoun was by far the most used in the depressive users dataset (80%), and the first- and the second-person plural pronouns were the least frequent (0.4% in both cases), this distribution being different from that of the control dataset (P<.001). Emotions related to sadness, anger, and disgust were more common in the depressive users and depressive tweets datasets, with significant differences when comparing these datasets with the control dataset (P<.001). As for negation words, they were detected in 34% and 46% of tweets in among depressive users and in depressive tweets, respectively, which are significantly different from the control dataset (P<.001). Negative polarity was more frequent in the depressive users (54%) and depressive tweets (65%) datasets than in the control dataset (43.5%; P<.001). CONCLUSIONS:Twitter users who are potentially suffering from depression modify the general characteristics of their language and the way they interact on social media. On the basis of these changes, these users can be monitored and supported, thus introducing new opportunities for studying depression and providing additional health care services to people with this disorder.

Dataset Information

TweepFake: About detecting deepfake tweets.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets