Deepfake images are getting harder to see. DeepGuard sees them.
DeepGuard is an image forensics tool built on a novel deepfake predictor (DFP) that combines VGG16 with a convolutional neural network to separate real faces from fake ones.
Fake media is a growing crime tool
Deepfakes swap a person's face or voice so convincingly that people and systems are fooled.
Mostly abusive
Over 96% of deepfakes are obscene content, according to a Sensity report cited in the study.
Real financial losses
In 2019 criminals used a faked CEO voice to request a $243,000 transfer.
Hard to detect
Digital forensics teams need reliable tools that keep up with new fakes.
A purpose-built deepfake predictor
DeepGuard pairs a proven image backbone with a custom CNN head trained on real and fake faces.
Correct calls on test images
- Hybrid VGG16 + CNN
- Trained on real and fake faces
- Single-image decisions
Few false alarms
- Fewer real images flagged
- Useful for triage
- Supports analyst review
Better than standard models
- Xception
- NAS-Net and MobileNet
- VGG16 alone
How DeepGuard works
From an uploaded image to a real-or-fake verdict.
- 1
Collect faces
A labelled dataset of real and deepfake face images trains the model.
DatasetLabels - 2
Extract features
VGG16 layers pull rich visual features from each face.
VGG16Transfer learning - 3
Classify
A CNN head decides whether the face is real or fake.
CNNDFP - 4
Report
The verdict supports investigators and content reviewers.
ForensicsTriage
What the research showed
Results from the DeepGuard study on a real-versus-fake face dataset.
| Model | Type | Outcome |
|---|---|---|
| DFP (VGG16 + CNN) | Proposed hybrid | Best: 94% accuracy, 95% precision |
| Xception | Transfer learning | Lower than DFP |
| NAS-Net | Transfer learning | Lower than DFP |
| MobileNet | Transfer learning | Lower than DFP |
| VGG16 | Transfer learning | Lower than DFP |
Source: Raza et al., Applied Sciences 12(19), 9820 (2022). doi:10.3390/app12199820
Who DeepGuard is for
Anyone who has to decide whether an image is real.
Cybersecurity teams
- Screen suspicious images
- Support fraud investigations
- Faster triage
Media and platforms
- Check user-uploaded content
- Limit fake news spread
- Protect public figures
Victims and legal teams
- Evidence that an image is fake
- Support takedown requests
- Counter blackmail
Where the evidence stands
DeepGuard is based on a published model tested on a research dataset.
Images only
The study covered still face images, not video or audio.
Research dataset
Performance on new generation methods needs ongoing testing.
Product work ahead
An upload tool, API and retraining pipeline are still to be built.
The team
Founded in 2022 by the research team behind the study.
Mubarak Almutairi
Co-author of the DeepGuard research.
Ali Raza
Co-author of the DeepGuard research.
Kashif Munir
Co-author of the DeepGuard research.
Frequently asked questions
What is a deepfake?
Synthetic media where a person's face or voice is replaced to look and sound real.
How accurate is DeepGuard?
The DFP model reached 94% accuracy and 95% precision in the published study.
Does it work on video?
The study tested images. Video frames could be checked one by one, but that still needs validation.
Who built it?
Researchers including Mubarak Almutairi.
Peer-reviewed science behind DeepGuard.
Read the full study behind the product.
Raza, A., Munir, K., & Almutairi, M. (2022). A Novel Deep Learning Approach for Deepfake Image Detection. Applied Sciences, 12(19), 9820.
https://doi.org/10.3390/app12199820