Artificial intelligence (AI) and deep-learning technologies have led to the development of so-called deepfakes. These are generated or manipulated video or audio recordings that can alter a person’s facial expressions, voice, or even their entire identity. While deepfakes have some legitimate uses in areas such as entertainment and art, their potential for misuse and for the spread of misinformation or damaging reputations has been recognised for several years. As the technology used to create deepfakes becomes more sophisticated, so there is a growing need to develop methods for deepfake detection.
Research described in the International Journal of Computational Science and Engineering introduces a hybrid deep-learning model that can itself improve the detection of deepfake content. The model combines two convolutional neural network (CNN) architectures, Inception ResNetV2 and Xception, along with long short-term memory (LSTM) networks. LSTM networks can process sequential data from video or audio segments and are particularly useful in spotting inconsistencies in manipulated media.
Shourya Chambial, Tanisha Pandey, Rishabh Budhia, and Balakrushna Tripathy of the Vellore Institute of Technology in Tamil Nadu, India, and Anurag Tripathy of Carnegie Mellon University in Pittsburgh, Pennsylvania, USA, trained their deepfake detector on a large dataset containing both real and manipulated video data. They were able to achieve an accuracy of almost 97 percent in tests. This suggests that the hybrid approach is capable of identifying subtle signs of manipulation in digital media and so decide whether a video is real or fake.
An important aspect of the research is that it emphasizes the importance of fine-tuning deep-learning models so that they work well with real-world data. An issue that commonly arises with certain types of AI model is that of “overfitting”, where the algorithm is too closely tied to the specific characteristics of its training data and struggles to perform on new, unseen data. In the current work, the team monitored performance and adjusted it to ensure it remained effective with a wide range of video content.
Chambial, S., Pandey, T., Budhia, R., Tripathy, B. and Tripathy, A. (2025) ‘Unlocking the potential of deepfake generation and detection with a hybrid approach’, Int. J. Computational Science and Engineering, Vol. 28, No. 2, pp.151–165.