MobileNetV2 Framework for Deepfake Detection and Real-Time Image Authentication: Safeguarding Users through an Accessible Web-Based System

Abstract

Deepfake technology causes serious threats to privacy, security, and trust in digital media. This paper presents a lightweight, real-time image authentication system based on MobileNetV2, designed to identify altered images through an easy-to-use web interface. The model achieved 80.38% test accuracy while maintaining low computational overhead, making it suitable for resource-constrained environments. Unlike traditional, resource-intensive solutions, the proposed approach prioritizes accessibility and practical deployment, aiming to effectively combat misinformation andidentity fraud.

Introduction

The rise of manipulated media, especially deepfake images created with advanced Generative Adversarial Networks (GANs), poses significant risks such as misinformation, privacy violations, and declining trust in visual content, especially on social media. To address these challenges, this paper proposes a lightweight, accessible, and real-time deepfake detection system based on MobileNetV2, a convolutional neural network optimized for efficient performance on web and edge devices.

The system allows users—technical or non-technical—to upload images via an easy-to-use web interface and instantly verify their authenticity without requiring extensive computational resources. The approach balances accuracy and speed, making it suitable for widespread use.

Key points:

• Background: Early detection relied on handcrafted features and shallow models, which were ineffective against sophisticated deepfakes. Deep learning models like CNNs, capsule networks, and architectures such as MesoNet and XceptionNet improved accuracy but often require heavy computation or struggle with generalization.

• MobileNetV2 Advantage: MobileNetV2 is lightweight, efficient, and suitable for real-time applications. It has been successfully applied in forgery and anomaly detection with good trade-offs between accuracy and resource use.

• Methodology: The model was trained on a Kaggle dataset containing 36,142 labeled real and fake images, with data augmentation and preprocessing steps to improve generalization. The final architecture uses a pretrained MobileNetV2 base with additional layers for binary classification (real/fake).

• Performance: The model achieved roughly 80% accuracy across training, validation, and test sets, with balanced precision and recall for both real and fake classes. It performs well in real-time detection scenarios and maintains responsiveness via the web interface.

• Web Interface: The system features a user-friendly online platform allowing instant image upload and authentication, bridging the gap between advanced AI detection and public usability. Images are classified and stored locally without reliance on external APIs or high-end hardware.

• Comparisons and Contributions: Compared to heavier models like XceptionNet (98% accuracy but resource-heavy) and lighter ones like MesoNet (84.3% accuracy), the proposed MobileNetV2 solution balances speed, accuracy, and deployment feasibility. Its novelty lies in combining effective detection with a practical, secure, and accessible web tool, supporting real-time deepfake image authentication.

• Future scope: While robust under standard conditions, the model could improve with expanded datasets, ensemble methods, and enhanced generalization to counter increasingly sophisticated deepfakes.

Conclusion

In this paper,MobileNetV2 Framework for Real-Time Image Authentication and Deepfake Detection is proposed and implemented. The solution combines a user-friendly web interface created with PHP and Python with a lightweight MobileNetV2 architecture optimized for deepfake detection. This paper prioritizes accessibility, deployability, and real-time performance in low-resource contexts, in contrast to conventional deep learning implementations that are limited to research labs or high-resource environments. This paper\'s distinctive feature is the smooth combination of an intuitive frontend with a deep learning backend, which allows uploaded deepfake images to be classified as either real or fake in real time. This fusion of deep-learning with web technologies bridges the gap between academic research and practical everyday use, an area often overlooked in existing literature. Additionally, the system\'s ability to provide both the classification result and the confidence score enhances interpretability and transparency, which will enable end users to make well-informed choices. This model demonstrated a great learning capacity and fair generalization over unseen data, with training accuracy of 81.43% and test accuracy of 80.38%.

References

[1] I. Goodfellow et al., “Generative Adversarial Nets,” Advances in Neural Information Processing Systems, vol. 27, 2014. [2] Y. Dong, J. Ding, and D. Zhou, “Visual misinformation detection via deep learning: A review,” IEEE Access, vol. 9, pp. 12302–12324, 2021. [3] Z. Zhao, D. Wang, T. Li, and H. Liu, “Detecting deepfake videos using recurrent neural networks,” in Proceedings of the IEEE ICASSP, 2020, pp. 2907–2911. [4] D. Afchar, V. Nozick, J. Yamagishi, and I. Echizen, “MesoNet: A Compact Facial Video Forgery Detection Network,” in Proc. WIFS, 2018. [5] H. H. Nguyen, J. Yamagishi, and I. Echizen, “Capsule-forensics: Using capsule networks to detect forged images and videos,” in ICASSP 2019 - IEEE International Conference on Acoustics, Speech and Signal Processing, 2019. [6] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “MobileNetV2: Inverted Residuals and Linear Bottlenecks,” in Proc. IEEE CVPR, 2018. [7] Z. Zhang, Z. Zhu, Y. Zheng, and K. Sun, “Face manipulation detection via feature disentanglement with low-rank regularization,” in Proc. ACM Multimedia, 2020. [8] J. H. Bappy, A. K. Roy-Chowdhury, and A. N. Mahmud, “Exploiting temporal inconsistencies for deepfake detection,” in Proc. CVPR Workshops, 2019. [9] A. Rössler, D. Cozzolino, L. Verdoliva, C. Riess, J. Thies, and M. Nießner, “FaceForensics++: Learning to detect manipulated facial images,” in Proc. ICCV, 2019. [10] Y. Li, X. Yang, P. Sun, H. Qi, and S. Lyu, “Celeb-DF: A Large-scale Challenging Dataset for DeepFake Forensics,” in Proc. IEEE CVPR, 2020. [11] M. Dolhansky et al., “The Deepfake Detection Challenge (DFDC) Preview Dataset,” arXiv preprint arXiv:1910.08854, 2019. [12] X. Chen, Y. Liu, W. Zhou, and T. Wang, “Edge-AI based framework for real-time deepfake detection,” IEEE Internet of Things Journal, vol. 9, no. 7, pp. 5306–5314, 2022. [13] F. Chollet, “On the measure of intelligence,” arXiv preprint arXiv:1911.01547, 2019. [14] Microsoft, “Video Authenticator Tool for Deepfake Detection,” [Online]. Available: https://news.microsoft.com [15] Deepware, “Deepfake Detection Scanner,” [Online]. Available: https://www.deepware.ai [16] F. Roffo, S. Melzi, and A. Vinciarelli, “XceptionNet for DeepFake Detection,” in Proc. IEEE International Conference on Image Processing (ICIP), 2020. [17] M. Tan and Q. Le, “EfficientNet: Rethinking model scaling for convolutional neural networks,” in Proc. ICML, 2019. [18] F. Iandola et al., “SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size,” arXiv preprint arXiv:1602.07360, 2016. [19] A. Agarwal, R. Farid, and S. Basu, “DetectDeep: A Deep Learning Pipeline for Detecting AI-generated Images,” ACM Digital Threats: Research and Practice, vol. 5, no. 1, 2021.

Copyright

Copyright © 2025 Mamabusetsa. E. Ntsau, D. Lalitha Bhaskari. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.