DeepFake Detection Using CNN Models with Explainable AI Techniques
Bipin P.R., Alen K Biju, M Balagopal, B. Shyam, Chamindu Adithya · 2025
The rapid advancements in artificial intelligence have enabled the creation of highly realistic, AI-generated fake images, commonly referred to as DeepFakes, which pose significant challenges to society by spreading misinformation and causing potential harm through social media platforms. This paper presents a robust DeepFake detection framework that combines high accuracy with enhanced interpretability. Multiple state-of-the-art deep learning models, including XceptionNet, EfficientNet, and MesoNet, are evaluated on the Celeb-DF v2 dataset, a benchmark for DeepFake detection, and an internal dataset containing diverse image manipulations. These models leverage unique convolutional neural network architectures, such as XceptionNet’s depthwise separable convolutions, EfficientNet’s compound scaling, and MesoNet’s lightweight design. To enhance transparency and trust in the detection process, Gradient-weighted Class Activation Mapping (Grad-CAM) is integrated, visualizing critical regions influencing predictions and providing actionable insights into model decisions. Experimental results demonstrate the framework’s effectiveness, with XceptionNet and EfficientNet achieving superior detection performance, while the integration of Grad-CAM ensures greater interpretability and accountability, making this approach a promising tool for mitigating disinformation risks on media platforms.