Robust deepfake detection through the AI-Guard mobile app for Real-Time image identification

Abstract

The proliferation of deepfake technologies has created an urgent need for robust, real-time detection systems capable of verifying image authenticity, particularly in mobile environments. This paper presents a scalable deepfake image detection framework integrated into the AI-Guard mobile application. Our approach leverages fine-tuned, computationally efficient convolutional neural network (CNN) architectures, including VGG19, InceptionV3, Xception, EfficientNetB0, ResNet50, and MobileNetV3Large. Trained on a large-scale, balanced dataset of over 450,000 real and fake images sourced from six publicly available datasets, the models incorporate advanced preprocessing, adversarial data augmentation, and optimized training pipelines. Among these, VGG19 achieved the highest generalization performance with 98.9% validation accuracy and 93.2% accuracy on previously unseen real-world data. The system supports real-time inference via a REST API, enabling practical mobile deployment with low latency. To support transparency and reproducibility, the curated training dataset has been made publicly available through our institutional repository. Our results demonstrate that AI-Guard offers an effective, deployable solution for forensic image verification, contributing to countering misinformation and enhancing trust in digital media.