We are happy to announce the new work by Fabrizio Falchi (ISTI CNR Pisa), Tiziano Fagni, Margherita Gambini, Maurizio Tesconi (IIT CNR) and Antonio Martella (Università di Trento) entitled “TweepFake: about Detecting Deepfake Tweets” published by Plos One.
The main aim of the work is the detection of automatically generated text from 23 bots mimicking 17 human users via several neural networks technologies (#GPT-2, #RNN, #LSTM, etc.) .
Here are the main takeaways:
- To the best of our knowledge TweepFake is the first analysis on “real” deep fake text on social media.
- The newest and more sophisticated generative methods (e.g., GPT-2) can produce high-quality short texts, difficult to unmask also for expert human annotators.
- The transformer-based language models provide very good word representations for fine-tuning based detection techniques.
- CHAR_GRU-based detector was the best at correctly labelling GPT2-tweets as bots.