Dubai Saraab Opens Deepfake Detection for Scrutiny

Dubai Electronic Security Center has announced Saraab, an AI model intended to detect manipulated video frame by frame and highlight suspect facial regions with a heatmap. Officials said at GISEC Global on 16 September that the tool would be released as open source, while a DESC executive cited 91% accuracy.

Key takeaways

Saraab is a government-developed detection model, not proof that any specific video is authentic or fake.
The 91% figure is an attributed claim; a public test set, class breakdown and operating threshold were not disclosed.
Open source could enable scrutiny, but the official repository and licence were not verified at package time.

Everyone else is reporting the accuracy headline; Lapaas Voice is defining the evidence teams need before relying on it.

What Saraab is designed to do

According to independent reporting from the event, Saraab analyses the full duration of a video rather than sampling only selected moments. It then uses a facial heatmap to indicate areas associated with suspected manipulation. DESC described the model as the first of its kind in the region and said an Emirati team developed it.

That workflow can help an analyst focus attention, but a heatmap is not a forensic verdict. Its meaning depends on preprocessing, model version, detection threshold and the types of manipulation represented in training and evaluation data. Compression, resizing and edits made by social platforms can also change a detector’s input.

A detector should supply a review signal inside an evidence process, not act as the final decision.

Why the 91% claim is incomplete

Accuracy compresses different mistakes into one number. For a fraud team, missed deepfakes and false alarms carry different costs. The claim cannot be compared meaningfully without the number and provenance of test videos, manipulation methods, demographic coverage, false-positive and false-negative rates, and whether the evaluation was independent.

The security context also raises handling questions. Sensitive videos may contain faces, voices, locations or confidential meetings. Organisations should verify whether analysis stays on their infrastructure, what telemetry is retained and whether uploaded files leave a controlled environment. Open-source availability can support that review, but only once the authentic publisher, code, model weights and licence are identifiable.

Saraab belongs beside defensive tools such as the Zscaler agentic SOC launch: both can prioritise signals, but neither removes accountable investigation.

What to verify at release

Teams should pin the official model version, checksum and licence; reproduce published metrics; and build a frozen test set reflecting their real video channels. Results should be logged with thresholds and analyst decisions. Until those materials appear, Saraab is a notable public-sector product announcement with a testable promise—not a validated universal deepfake detector.

Item	Verified detail
Announced	16 September 2026
Developer	Dubai Electronic Security Center
Purpose	Deepfake-video detection
Method described	Frame-by-frame analysis and facial heatmap
Accuracy claim	91%, attributed to DESC
Availability	Open-source release promised; repository not yet verified

Related Lapaas Voice coverage: related technology coverage and related technology coverage.

Frequently asked questions

What is Saraab?

Saraab is a Dubai Electronic Security Center AI model announced for analysing video frames and highlighting possible facial manipulation.

Is Saraab available to download?

DESC said it would be open source, but no official public repository or licence was verified at package time.

Does 91% accuracy prove it will catch every deepfake?

No. The figure lacks a public test-set and threshold breakdown, so organisations should validate the model on representative material and preserve human review.