XAI-Driven Deep Neural Network Approach for Synthetic Image Identification

Le résumé fourni par la source

The wide availability of deepfake generation models has led to the amalgamation of fake media with the original ones. It raises questions about the validity of images on the internet. Deepfakes pose threats to society, including financial losses. Traditional and machine learning detection measures are infeasible due to the inability to detect them efficiently. This led to a shift to Deep Learning (DL) approaches. However, the black-box nature of their prediction inhibits their real-life deployment. To address it, this study proposes a custom convolutional neural network (CNN) architecture to detect deepfake images and provide explainability of decisions. Gradient-weighted Class Activation Mapping (Grad-CAM) is used to visualize the region-of-interest that affects decision-making the most. The classifier is organized into four convolutional blocks of 64,128,256, and 512 filters, respectively. A residual connection is added to the final convolutional block to retain vital feature representations. Three fully connected layers are used for classification in 512, 128, and 2 neurons. Various callbacks and training optimization mechanisms ensure a smooth training process without overfitting. The 140K real and fake faces dataset is used to train and evaluate the model. It achieved a test accuracy and AUC of 98.5% and 0.9987, respectively. Various other near-ideal parameters with minimal error rate demonstrate the efficacy of the model. This model presents a computationally balanced architecture to detect deepfake images efficiently. Its domain-agnostic design ensures easy adoption.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.