A Comprehensive Review of EfficientNet-Lite and Attention-Driven Deepfake Detection for Real-Time Video Streams

Abstract

Deepfake technology, powered by advanced AI mod- els like GANs, poses serious threats by generating hyper-realistic fake videos, which are increasingly used in misinformation, fraud, and cybercrime. Despite rapid progress in deepfake detection methods, challenges remain in achieving high accuracy with real- time processing and robustness to evolving synthesis techniques. Recent works have leveraged various deep learning architectures, such as CNNs, LSTMs, Transformers, and hybrid models, with accuracy ranging from 85 to 98 percent on benchmark datasets. For example, EfficientNet-based models combined with attention mechanisms have shown promising detection accuracy above 95 percent, while multimodal approaches incorporating audio and visual features further improve robustness. However, current so- lutions often struggle with computational efficiency and adapting to new deepfake generation methods in live video streams. This survey highlights these research gaps and critically assesses state- of-the-art AI-based real-time deepfake detection systems, includ- ing techniques using corneal reflection analysis, active forensic probing, and multi-scale attention networks. We conclude by outlining future directions focused on improving detection speed, adversarial robustness, and deploying practical solutions for live video surveillance and conferencing environments.

Introduction

The rapid advancement of deepfake technology, powered by AI and generative adversarial networks (GANs), poses serious threats to visual media integrity, public trust, and societal security. Deepfakes have evolved from simple face swaps to highly realistic synthetic videos, enabling political manipulation, financial fraud, identity theft, harassment, and psychological harm. High-profile incidents, such as deepfake-based scams and fabricated political videos, illustrate the technology’s wide-reaching risks.

Detecting deepfakes has become a major research focus. Early methods relied on visual forensics, analyzing lighting, facial expressions, and physiological cues, but these proved inadequate against high-quality fakes. Modern detection approaches leverage AI, combining spatial feature extraction (via CNNs like EfficientNet) with temporal modeling (RNNs, LSTMs, GRUs, and temporal convolutional networks) to capture frame-to-frame inconsistencies. Hybrid architectures integrating spatial and temporal analysis have shown superior performance. Active forensic methods, such as corneal reflection analysis in video conferencing, introduce proactive verification but face practical limitations.

Challenges in deepfake detection include cross-dataset generalization, adaptation to new generation techniques (like diffusion-based models), computational costs, and real-time deployment. Edge computing, model quantization, and pruning are explored to reduce latency and enable lightweight deployment.

Surveyed systems range from CNN-only frame-level detectors to hybrid spatiotemporal architectures and active forensic approaches. Key trends include:

• EfficientNet-based spatial feature extraction for subtle artifact detection.

• Temporal modeling to identify sequential inconsistencies.

• Lightweight mobile and real-time frameworks for streaming applications.

• Multi-modal approaches integrating audio-visual signals.

Despite progress, limitations remain, including dataset overfitting, vulnerability to adversarial attacks, poor generalization across manipulation types, and high computational requirements. Overall, the field emphasizes a balance between detection accuracy, efficiency, and practical deployability, with ongoing research needed to address evolving deepfake generation techniques.

Conclusion

In summary, deepfake detection has rapidly evolved from spatial-only CNN classifiers to sophisticated hybrid architec- tures that integrate temporal modeling, active forensics, and scalable deployment pipelines; however, significant gaps re- main in cross-dataset generalization, long-range temporal rea- soning, low-latency mobile inference, robust active authentica- tion under diverse conditions, and multimodal fusion. Bridging these gaps will require novel domain adaptation strategies, lightweight yet accurate architectures, privacy-preserving chal- lenge–response protocols, integration of audio and physiolog- ical signals, and standardized evaluation metrics that reflect real-world constraints. Only through such interdisciplinary efforts can next-generation systems reliably safeguard digital media integrity in an increasingly adversarial landscape.

References

[1] R. Khaled, H. M. Moftah, F. K. Alsheref, A. S. Assiri, K. H. Rahouma, and M. Kayed, ”Boosting Deepfake Detection Accuracy with Unsharp Masking and EfficientNet Models,” International Journal of Advanced Computer Science and Applications (IJACSA), vol. 16, no. 8, pp. 322- 333, 2025. [2] S. Chaudhauri, V. R. Goudar, S. Hazam, and V. Ranjan, ”AI-Powered Real-Time Deepfake Detection,” International Journal of Research Pub- lication and Reviews, vol. 6, no. 4, pp. 6577-6581, April 2025 [3] S. Geetha, V. S. Murali, V. G. K, and P. T. S, ”Deepfake Detection using TCN and EfficientNet-B3,” International Journal on Science and Technology (IJSAT), vol. 16, no. 2, pp. 1-12, April-June 2025 [4] K. V. A. Reddy, Lochan S, Shrusthi, E. P. Goud, and M. Swapna, ”Development of AI/ML-Based Solution for Detection of Face-Swap Deep Fake Videos,” International Journal of Scientific Research and Engineering Trends, vol. 11, no. 2, pp. 2689-2696, Mar-Apr 2025. [5] H. Guo, X. Wang, and S. Lyu, ”Detection of Real-Time Deepfakes in Video Conferencing with Active Probing and Corneal Reflection,” arXiv preprint arXiv:2210.14153v1, Oct. 2022 [6] Paravision Inc., ”A Practical Guide to Deepfake Detection: Definitions, [7] Challenges, and Solutions,” Technical Report, October 2024 [8] M. Arkachari, U. Sadalage, O. Naik, P. Chandake, and P. Sonnad, ”The State of Live Deepfake Detection in Streaming Platforms,” International Journal of Advanced Research in Science, Communication and Tech- nology (IJARSCT), vol. 5, no. 5, pp. 498-503, June 2025. [9] Abhinav, S. Singh, and A. Bharadwaj, ”Realtime Deep-fake Detection on Video Streams,” International Journal of Research Publication and Reviews, vol. 6, no. 4, pp. 7809-7811, April 2025. [10] G. Sangar and V. Rajasekar, ”Optimized classification of potato leaf disease using EfficientNet-LITE and KE-SVM in diverse environments,” Frontiers in Plant Science, vol. 16, article 1499909, 2025. DOI: 10.3389/fpls.2025.1499909. [1] G. Harshit, M. C. Reddy, and Y. V. S. Sarath, ”Deepfake Detection using Efficientnet-B0 and GRU,” International Journal of Engineering Research and Technology (IJERT), vol. 14, no. 4, pp. 1-8, April 2025. [2] C. Sahu, ”Deepfake Detection System,” International Journal of Engi- neering Research and Technology (IJERT), vol. 14, no. 5, pp. 1-12, May 2025. [3] M. Alrashoud, ”Deepfake video detection methods, approaches, and challenges,” Journal of Computer and Information Sciences, vol. XX, no. XX, pp. XX-XX, 2025. [4] R. Sunil, P. Mer, A. Diwan, R. Mahadeva, and A. Sharma, ”Exploring autonomous methods for deepfake detection: A detailed survey on techniques and evaluation,” International Journal of Computer Science and Engineering, vol. XX, no. XX, pp. XX-XX, 2025. [5] G. N. Vivekananda, T. R. Mahesh, M. Gupta, A. Thakur, and A. Sayal, ”Refining digital security with EfficientNetV2-B2 deepfake detection techniques,” Journal of Cybersecurity and Digital Forensics, vol. XX, no. XX, pp. XX-XX, 2025. [6] S. S. Chorage, A. Barabde, J. Balsaraf, and S. Batwal, ”Deepfake Detection Using Deep Learning,” International Research Journal on Advanced Engineering Hub, vol. 3, no. 8, pp. 3427-3431, Aug. 2025.

Copyright

Copyright © 2025 Rakhi Meshram, Suyog Talape, Manasi Thorat, Rutik Vidhate. 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.