A Deepfake Detection System for Law Enforcement Applications Using Xception

Abstract Deepfake imagery highly realistic synthetic media generated by AI poses a growing threat to digital evidence and public trust. We present a deep learning–based system for law enforcement to detect deepfakes in suspect images or videos. Authorized officers submit content via a secure API; the system processes each input with a convolutional neural network (based on Xception) that classifies media as authentic or manipulated. The detector is trained on a balanced dataset of real face images (from public sources such as the FFHQ dataset) and synthetic faces generated by StyleGAN3. StyleGAN3’s alias-free generator produces high-quality fakes that improve classifier robustness. In testing, our model attains high accuracy (~93%), with precision and recall above 90%, and a low false-positive rate, indicating reliable discrimination between real and fake faces. By automating media verification, this system helps address the urgent need for advanced deepfake detection in criminal investigations.