Abstract

This paper offers a comprehensive framework for real or fake image classification based on three classifiers: a Standard Convolutional Neural Network (CNN), an EfficientNetV2 model based on transfer learning, and a re-trained GAN discriminator to address the challenges in deepfake detection. The CNN, with four convolutional blocks and dropout regularization, offers computational efficiency (87.2% accuracy, 15 ms/image inference), while EfficientNetV2 utilizes pre-trained ImageNet weights to achieve state-of-the-art performance (94.7% ac-curacy, AUC: 0.98) using hierarchical feature extraction. The fine-tuned and adversarial-pretrained GAN discriminator demonstrates niche strength in the detection of synthetic artifacts (91% recall for GAN-generated fakes). Training used augmented sets (rotation, shifts, and shear) to increase the generalization boost, with loss optimization and early stopping (binary cross-entropy) controlled through validation. Normalized test set validation affirmed EfficientNetV2's capability at balancing recall (94%) with precision (95%), although the GAN discriminator recorded a lead in adversarial resilience. All the models blended, an ensemble model achieved maximum accuracy (96.1%), under complementarities. Computational baselines showed trade-offs EfficientNetV2 accu-racy vs. resource bias (2.5-hour training), the CNN edge-compatibility, and the GAN discriminator arti-fact-sensitive specialization. The work encourages hybrid architectures and ensemble approaches to balance out single-model vulnerabilities, offering a flexible toolkit for deepfake warfare while emphasizing the need for hardware-aware deployment techniques and ongoing adaptation to changing synthetic approaches.