Efficientnet-Based Deep Learning Approach for Video Forgery Detection and Authentication
V Gowri Priyaa, M Jaya Harrish, M Udhayakumar, N Jothieswaran, Kalluri sai Dinesh · 2024
The technological advancements in artificial intelligence have made it a lot easier to create forged videos that are difficult to distinguish from reality. Fake videos also called deep fakes are created with greater accuracy and precision. Detecting and removing fake data on the internet can prevent misinformation and rumors from spreading. To achieve this, detection methods must be robust, generalized, fast, and accurate enough to detect fake data. In this paper, deepfakes are created using Generative Adversarial Network (GAN) and used for dataset training. The deepfakes are found to be different from real ones by various parameters like facial expressions, irregularities in the image, etc. This project focuses on detecting the manipulated face of the person in a frame using the EfficientNet B4 algorithm. The EfficientNet B4 model is more accurate than EfficientNet B0 and less complex than EfficientNet B7. The modified EfficientNet B4 outperforms the existing EfficientNet B0 in terms of accuracy. The probabilities of deep fakes in each frame are calculated and on average the video is detected as real or fake. This model demonstrates a very successful detection rate of more than 92%. Finally, modified EfficientNet B4 is compared with other models’ performance.