LPP-GAN: A Mask-Guided Generative Framework for Privacy- Preserving Face Anonymization on Mobile Devices

Abstract

With the fast growth of smart phone photography and mobile content sharing, privacy protection becomes increasingly important. While generative adversarial networks (GANs) are promising solutions to realistic face anonymization, current methods are mostly tested in controlled desktop or emulator environments and practical mobile deployment is largely unexplored. We propose in this work a light-weight mask-guided LPP-GAN framework for real-device face anonymization on Android smartphones. We train the model at 256×256 resolution using a combined loss strategy that balances visual realism, background preservation, and effective identity obfuscation. The trained generator is exported in ONNX format for deployment. It is integrated in an Android application using ONNX Runtime with CPU-only execution. Experimental results on 175 test images show that the proposed method obtains 32.13 dB PSNR and 0.96 SSIM without any detectable facial identity (0.0% detection rate). Real-device benchmarking over 20 images with 1-3 faces reports a mean latency of 1380.16 ms (95% confidence interval: 1224.79-1535.53 ms) and 0.82 fps. The results illustrate the performance gap between emulator-based testing and physical hardware validation and underscore the importance of real-device evaluation for edge privacy-preserving AI systems