The rapid advancement of synthetic image gener- ation through generative adversarial networks (GANs), autoen- coders, and diffusion models has significantly increased the diffi- culty of distinguishing authentic visual content from manipulated or AI-generated imagery. This paper presents a deep learning- based three-class classification framework capable of separating AI-generated, deepfake, and real images using an EfficientNet-B0 backbone. A series of targeted optimization strategies—including label smoothing, warmup–cosine learning rate scheduling, mixed- precision training, and staged unfreezing—are integrated to enhance model stability and generalization. Using a dataset of 9,999 RGB face images sourced from the HuggingFace repository, the proposed system achieves a validation accuracy of 99.8%, supported by strong ROC–AUC performance and minimal overfitting across epochs. The experimental results demonstrate that lightweight yet well-regularized convolutional architectures remain highly effective for modern image forensics. The framework thus provides a practical foundation for future work in manipulation localization, multimodal forensics, and domain-adaptive detection.
Keywords
Deepfake detection, AI-generated images, CNN, EfficientNet-B0, image forensics, transfer learning.
Conclusion
This study presents a robust and efficient multi-class CNN framework capable of distinguishing AI-generated, deepfake, and real facial images with a validation accuracy of 99.8%. By integrating EfficientNet-B0 with compound scaling, warmup– cosine learning rate scheduling, label smoothing, mixed- precision optimization, and staged unfreezing, the proposed approach demonstrates exceptional stability, rapid conver- gence, and strong generalization across heterogeneous genera- tive sources. The near-diagonal confusion matrix and smoothly converging accuracy and loss curves highlight the model’s ability to capture subtle yet distinct artifact signatures across synthesis modalities.
The findings underscore the impact of well-engineered training pipelines when addressing modern synthetic imagery, where generative models increasingly minimize detectable artifacts. Efficient feature extraction, carefully constructed augmentations, and progressive fine-tuning collectively play a decisive role in achieving high discriminative power in multi- class forensic settings.
Despite its strong performance, the current work opens several promising avenues for further exploration. Future di- rections include:
•
Domain Adaptation: Enhancing resilience to unseen distributions, compression settings, and cross-platform generative pipelines.
•Video and Multimodal Deepfake Forensics: Extending detection to temporal, audio, and physiological modalities for comprehensive multimedia analysis.
•Manipulation Localization: Augmenting the classifier with spatial prediction modules capable of highlighting manipulated regions.
•Adversarial Robustness: Investigating vulnerabilities posed by adversarial perturbations and designing coun- termeasures for real-world deployment.
•Lightweight and Edge-Friendly Models: Optimizing model architectures for forensic applications on mobile and embedded devices.
In summary, this work establishes an effective, high- performing baseline for multi-class synthetic image forensics and contributes meaningful insights for the development of next-generation detection systems in an era of rapidly advanc- ing generative technologies.
References
[1]F. Marra, D. Gragnaniello, L. Verdoliva, and G. Poggi, “Do gans leave artificial fingerprints?” IEEE Transactions on Information Forensics and Security, vol. 14, no. 11, pp. 2756–2769, 2019.
[2]R. Corvi, M. Kettunen, and B. Boehm, “Detection of diffusion model generated images using frequency-domain analysis,” Forensic Science International: Digital Investigation, vol. 43, p. 301460, 2022.
[3]L. Verdoliva, “Media forensics and deepfakes: A survey,” IEEE Journal of Selected Topics in Signal Processing, vol. 14, no. 5, pp. 910–932, 2020.
[4]S. Wang, X. Chen, J. Yang, and W. Liu, “Cnn-based image forensics: A comprehensive study,” ACM Computing Surveys, vol. 53, no. 6, pp. 1–36, 2020.
[5]Y. Hamid, S. Elyassami, Y. Gulzar, V. R. Balasaraswathi, T. Habuza, and S. Wani, “An improvised cnn model for fake image detection,” International Journal of Information Technology, vol. 15, no. 1, pp. 5–15, 2023.
[6]Y. Patel, S. Tanwar, P. Bhattacharya, R. Gupta, T. Alsuwian, I. Davidson, and T. F. Mazibuko, “An improved dense cnn architecture for deepfake image detection,” IEEE Access, vol. 11, pp. 22 081–22 095, 2023.
[7]Y. Qian et al., “Thinking in frequency: Defense against deepfake detection via frequency-domain augmentation,” ECCV, 2020.
[8]C. C. Hsu, Y. X. Zhuang, and C. Y. Lee, “Deep fake image detection based on pairwise learning,” Applied Sciences, vol. 10, no. 1, p. 370, 2020.
[9]Y. Chai, H. Wang, and Q. Li, “Efficientnet-based deepfake detection using improved transfer learning,” IEEE Access, vol. 8, pp. 223 854– 223 865, 2020.
[10]B. R. Barik, A. Nayak, A. Biswal, and N. Padhy, “Practical evaluation and performance analysis for deepfake detection using advanced ai models,” Engineering Proceedings, vol. 87, no. 1, p. 36, 2025.
[11]D. Samal, P. Agrawal, and V. Madaan, “Deepfake image detection and classification using conv2d neural networks,” in Proc. International Workshop on Computational Intelligence (ICAIDS), 2023, pp. 1–7.
[12]O. Singh, S. Patel, and S. Singh, “Deepfake detection of images using deep learning techniques,” International Journal of Creative Research Thoughts, vol. 12, no. 3, 2024.
[13]Y. Li et al., “Celeb-df: A large-scale challenging dataset for deepfake forensics,” in CVPR, 2020, pp. 3207–3216.
[14]B. Dolhansky et al., “The deepfake detection challenge (dfdc),”
arXiv:2006.07397, 2020.
[15]B. Zi et al., “Wilddeepfake: A challenging real-world dataset for deepfake detection,” in ACM Multimedia, 2020, pp. 2023–2032.
A. Ro¨ssler et al., “Faceforensics++: Learning to detect manipulated facial images,” in ICCV, 2019, pp. 1–11.
How to Cite This Paper
Shashank Mani Tripathi, Priyanshu Srivastava, Devansh Yadav, Prachi Verma (2025). Classification of AI-Generated, Deepfake and Real Images Using CNN. International Journal of Computer Techniques, 12(6). ISSN: 2394-2231.
