Unmasking Deepfakes: A Multi-Modal Hybrid Detection Framework

The rapid evolution of deepfake technology has intensified the challenge of ensuring media authenticity, driving the need for sophisticated detection methods that go beyond conventional techniques in both adaptability and effectiveness. This paper introduces an innovative deepfake detection system that seamlessly integrates behavioral and visual analysis, eliminating the dependency on custom training datasets. By harnessing pre-trained models from trusted repositories, the system employs a triple-detection pipeline — comprising MTCNN, DLib, and Mediapipe Face Mesh — to achieve reliable face and landmark identification across a wide range of video inputs, ensuring resilience even with challenging or low-quality footage. At its core, the framework analyzes a rich set of features to distinguish authentic from synthetic content, including eye blinking dynamics such as frequency, period, duration, and symmetry, alongside lip movement consistency and temporal frame coherence assessed through optical flow. These behavioral cues are complemented by EfficientNet-B7, a state-of-the-art model that enhances detection by identifying pixel-level anomalies often present in deepfake videos. Implemented on Google Colab, this system processes user-uploaded videos in real-time, employing an optimized ensemble method with confidence-weighted scoring to classify content as "Real" or "Fake," offering a practical and accessible solution for media verification. Extensive debugging and adaptive threshold tuning bolster the system’s reliability against modern deepfakes, addressing the shortcomings of single-feature approaches like the original DeepVision framework. By combining multiple detection modalities and leveraging cloud-based computation, this lightweight and scalable tool surpasses traditional limitations, providing a robust defense against synthetic media. This work represents a significant step forward in deepfake detection, adaptable to the evolving landscape of digital content manipulation and suitable for real-world applications