Detecting Deepfake: Using Image Similarity Parameter

The rise in the deepfake media—digitally altered media (images, audio files, videos, etc.) has posed a moral threat to the human race. As deepfake generators continue to rise very alarmingly, detection methods cannot keep up with the pace. This work aims to use different similarity measures between two images to quantify the similarity score between the real image directory and the altered image. The proposed model is trained on a small dataset of 600+ images with a limited cloud Graphics processing unit (GPU). Deepfake and deepswap images were generated by using open packages (ROOP). Different similarity parameters such as Structural similarity index measure (SSIM), Multi-scale structural similarity index measure (MS-SSIM), Mean square error (MSE), and Euclidean distance are used along with a CNN-based feature extraction method, 2D convolution layer (Conv2D). This work compares two different algorithms, decision tree and random forest, to determine our dataset’s accuracy and best-fit model. Evaluation metrics such as—accuracy, Area under the Receiver Operating Characteristic Curve (AUC-ROC), and F1 score were used to find the best-fitting model. The two classification models, Random Forest and Decision Tree, were deployed, and both provided a good result, but the Random Forest turned out to be the best fit for our data. It provided an accuracy of 79.53 and 82.23%, whereas the Decision Tree model had 74.13 and 73.74% accuracy for the two different target deepfakes generated by ROOP and Face_Recognisation, respectively.