Deepfake Image Detection Using Yolov8

Abstract

Due to growing number of fake media and the possible problems with misinformation and identity theft, deepfake image recognition has become a hot topic. In order to evaluate the efficacy of widely recognized deep learning models in detecting and classifying deepfake images, we compare and contrast YOLO (You Only Look Once) V8, CNN (Convolutional Neural Network), combination of LSTM (Long Short-Term Memory) with CNN and ResNet in this paper. The capacity to recognize photos that have been manipulated effectively is essential for maintaining trust regarding digital media and preventing the spread of deceptive data. The objective of our research is to explain each model's abilities, limitations and performance in the context of deepfake image recognition. We evaluated each model's accuracy, precision, and F1-scores using a dataset of 190402 images, half real and half fake that cover a wide spectrum of deep fake images. Based on the evaluation metrics, each model was ranked in terms of its ability to separately describe and determine original versus generated photos with focus on detecting subtle changes eluding human senses. The comparison research shows each model's nuanced capabilities, offering light on the implications for applications in the real world like detecting and controlling the spread of deep fake information across numerous internet platforms. Our findings help towards building deep fake detection systems, further understanding the comparative performance of state of the art deep architectures. This study shall inform more effective deep fake detection systems that can guarantee trust and security in digital media environments. Accuracy achieved by model is 96%, precision score was 94%, recall value is 0.98, F1 score has a value of 0.95.

Topics & Concepts