The New Paradigm of Deepfake Detection at the Text Level

Applied Sciences, 2025, Vol 15, n°5 https://doi.org/10.3390/app15052560

Autor: 

1. Cosmina-Mihaela Rosca, Department of Automatic Control, Computers and Electronics, Faculty of Mechanical and Electrical. Engineering, Petroleum-Gas University of Ploiesti, 39 Bucharest Avenue, 100680 Ploiesti, Romania; cosmina.rosca@upg-ploiesti.ro;  https://orcid.org/0000-0003-0827-3321
2. Adrian Stancu, Department of Business Administration, Faculty of Economic Sciences, Petroleum-Gas University of Ploiesti, 39 Bucharest Avenue, 100680 Ploiesti, Romania; astancu@upg-ploiesti.ro; https://orcid.org/0000-0002-5366-8149
3. Emilian Marian Iovanovici; Department of Automatic Control, Computers and Electronics, Faculty of Mechanical and Electrical Engineering, Petroleum-Gas University of Ploiesti, 39 Bucharest Avenue, 100680 Ploiesti, Romania; emilian.iovanovici@student.upg-ploiesti.ro; https://orcid.org/0009-0007-5746-8826 Personal 

Keywords: text authenticity; deepfake texts; AI-generated text; written style model; text manipulation; ML model; author identification.

Abstract: The world is currently facing the issue of text authenticity in different areas. The implications of generated text can raise concerns about manipulation. When a photo of a celebrity is posted alongside an impactful message, it can generate outrage, hatred, or other manipulative beliefs. Numerous artificial intelligence tools use different techniques to determine whether a text is artificial intelligence-generated or authentic. However, these tools fail to accurately determine cases in which a text is written by a person who uses patterns specific to artificial intelligence tools. For these reasons, this article presents a new approach to the issue of deepfake texts. The authors propose methods to determine whether a text is associated with a specific person by using specific written patterns. Each person has their own written style, which can be identified in the average number of words, the average length of the words, the ratios of unique words, and the sentiments expressed in the sentences. These features are used to develop a custom-made written-style machine learning model named the custom deepfake text model. The model’s results show an accuracy of 99%, a precision of 97.83%, and a recall of 90%. A second model, the anomaly deepfake text model, determines whether the text is associated with a specific author. For this model, an attempt was made to determine anomalies at the level of textual characteristics that are assumed to be associated with particular patterns of a certain author. The results show an accuracy of 88.9%, a precision of 100%, and a recall of 89.9%. The findings outline the possibility of using the model to determine if a text is associated with a certain author. The paper positions itself as a starting point for identifying deepfakes at the text level.

Idioma: Inglés. 

Publicación: 27 de Febrero 2025.

Volumen: Applied Science 2025 Vol 15, n°5. 

Enlace permanente.