TweepFake: About detecting deepfake tweets.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computational Social Science

Background:

• Advanced language models like GPT-2 enable sophisticated AI to generate human-like text.
• Malicious actors can use these models to create deepfake messages for social bots, potentially disrupting public discourse.
• Existing research lacks datasets and methods for detecting machine-generated text specifically on social media platforms.

Purpose of the Study:

• To introduce TweepFake, the first dataset of authentic deepfake tweets posted on Twitter.
• To provide a benchmark for evaluating deepfake text detection methods in a social media context.
• To stimulate research into identifying AI-generated content on social networks.

Main Methods:

• Collected 25,572 tweets from 23 bots employing various generation techniques (Markov Chains, RNN, LSTM, GPT-2) and 17 imitated human accounts.
• Ensured a balanced dataset with an equal number of human-generated and bot-generated tweets.
• Evaluated 13 state-of-the-art deepfake text detection methods on the TweepFake dataset.

Main Results:

• The TweepFake dataset presents a significant challenge for current deepfake detection techniques.
• Established a baseline performance for 13 different detection methods on real-world deepfake social media data.
• Demonstrated the feasibility of creating and utilizing a dataset of actual deepfake tweets.

Conclusions:

• TweepFake serves as a vital resource for advancing research in social media deepfake detection.
• The dataset and baseline evaluations highlight the need for more robust detection systems against AI-generated disinformation.
• Further development is encouraged to address the challenges posed by sophisticated machine-generated text on social platforms.