Deepfakes are manipulated media that modify videos or images to misrepresent individuals or events. Deepfake detection system utilizing three distinct deep learning approaches: a customized ResNet model, MobileNetV2, and ResNet50. The study focuses on accurately identifying manipulated images, particularly those involving human faces, by comparing the effectiveness of these different architectures. The system employs a dataset from Kaggle containing both real and fake images, with the implementation of custom preprocessing techniques and data augmentation methods. The customized ResNet model features modified architecture with skip connections and residual blocks, while MobileNetV2 offers computational efficiency and ResNet50 provides robust feature extraction capabilities. The research includes the development of a user interface that enables quick image analysis, providing instant feedback on image authenticity with confidence scores. Comparative analysis revealed that the customized ResNet architecture achieved superior performance in detecting manipulation artifacts compared to the pre-trained models. The system incorporates multiple modules including data collection, preprocessing, detection, and analytics, making it a complete solution for deepfake detection. This research contributes to the field of digital media authentication by demonstrating that tailored deep learning architectures can provide more accurate and reliable detection while maintaining practical usability through an intuitive interface.