The Core AI — EfficientNet-B4
At the heart of DeepScan AI is a deep learning model called EfficientNet-B4. This model was originally trained to recognize thousands of different objects and scenes from millions of images. It has learned to understand visual patterns at a very deep level — textures, shapes, lighting, and fine details that the human eye might miss.
DeepScan AI uses this model as a "visual fingerprinting" tool. For each face in the video, the model creates a unique numerical description of what it sees. By comparing these descriptions across frames, the system can tell whether the face is behaving consistently — like a real person would — or jumping around in ways that suggest AI generation.
Why This Approach Works
When a deepfake is created, an AI generates each frame of the face somewhat independently. Even though the result looks smooth to the human eye, the underlying visual patterns are slightly different from frame to frame in ways that are hard to control. EfficientNet-B4 is sensitive enough to pick up on these subtle inconsistencies.
Real videos don't have this problem. A real person's face changes naturally and smoothly between frames — the visual fingerprints stay consistent throughout the video.
The Supporting Checks
The AI model alone isn't enough — deepfakes come in many forms. That's why four additional computer vision checks run alongside it:
Sharpness Analysis
Checks whether the face stays consistently sharp across all frames. Deepfake generators sometimes produce frames with uneven sharpness that real cameras don't.
Edge Detection
Looks for unnatural sharp edges at the boundary of the face — a common sign that a face has been digitally pasted onto a different body or background.
Color Analysis
Compares the color and lighting of the face against the rest of the scene. A face taken from a different video often has slightly different color tones that don't match the background.
Motion Analysis
Measures how smoothly the face moves between frames. Deepfakes can produce subtle flickering or unnatural movement patterns that real faces don't exhibit.
How the Final Decision is Made
All five checks produce individual scores. These scores are combined — with the AI model's result carrying the most weight since it's the most reliable signal. The combined score is compared against a threshold to produce the final REAL or FAKE verdict.
The confidence percentage shows how far the score is from the decision boundary. A 90% confidence REAL result means the video passed all checks by a wide margin. A 75% confidence FAKE result means some signals were detected but not all were strongly triggered.
Limitations
No detection system is perfect. DeepScan AI works best on face-swap deepfakes where a person's face has been replaced. It may be less effective on:
- Very short videos (under 2 seconds) — not enough frames to analyze
- Very low quality or heavily compressed videos
- Videos with no visible face
- Highly sophisticated AI-generated content from the latest generation tools
Always use the results as one piece of evidence alongside other context, not as a definitive judgment.