Verify the Evidence Without Exposing the People Inside It
Journalists, NGOs, fact-checkers and public-interest investigators regularly receive photos and videos that may be important but reviewing the raw material can expose the people inside it before anyone has decided whether it is relevant or safe to publish.
A short video may reveal a vulnerable subject, a bystander or a vehicle licence plate. Raw files may be opened, forwarded or uploaded during review, creating unnecessary privacy exposure before an editorial decision has been made.
The same material may later circulate as a crop, screenshot, recompressed copy or screen recording. These transformations can separate the content from the source information recorded when it was first collected and cause teams to review different versions of the same material repeatedly.
Privacy-Gate is a local, privacy-first visual evidence review system that creates a protected working copy before sensitive footage enters routine review or distribution.
It also preserves supplied source context and uses Metis-powered visual retrieval to reconnect incoming media with related material already held in a controlled local archive.
Verify the evidence without exposing the people inside it.
The Solution
Privacy-Gate processes images and short pre-recorded videos locally before they enter the normal review and sharing workflow.
Axelera Metis performs the main visual inference workloads:
- detecting faces;
- proposing licence-plate regions;
- extracting visual features used to identify related archive items.
The system then generates a protected reviewer copy in which sensitive regions can be blurred or covered.
The reviewer remains in control. They can inspect proposed masks, correct missed regions, reject incorrect detections and approve the final protected export.
This creates a practical division of responsibility:
Metis accelerates first-pass visual analysis. Privacy-Gate gives the reviewer control over the final privacy decision.
Preserving Source Context Through Visual Retrieval
When public-interest media is collected, Privacy-Gate records the supplied public-source reference, collection time and relevant source notes alongside the local archive record.
The system also creates visual representations of the incoming media. For video, selected keyframes are processed on Metis to generate compact visual embeddings.
When a new item enters Privacy-Gate, its hashes and visual embeddings are compared with the controlled local archive. This allows the system to surface:
- an exact copy that was previously archived;
- a recompressed version of an earlier item;
- a crop or screenshot derived from archived media;
- the source reference previously recorded for the related item;
- a possible parent-derivative relationship for reviewer confirmation.
For example, a reviewer may receive a cropped screenshot with no useful metadata. Privacy-Gate can compare its visual content with earlier archive records and surface the likely related original together with the public-source reference that was recorded when that original was collected.
This does not require internet-wide crawling. Privacy-Gate preserves known source context and helps reconnect later derivatives to that context through local, Metis-powered visual retrieval.
A deliberately selected hard negative, an unrelated item with similar visual content, will also be included in the evaluation to measure whether the retrieval system can avoid misleading matches.
Why Axelera Metis
Metis is not an accessory to the application. It performs the inference-heavy stages that make both sides of Privacy-Gate possible.
Privacy Protection
Face and licence-plate detection must be applied across many video frames. Running these models locally on Metis allows Privacy-Gate to create a protected reviewer copy without first sending sensitive raw footage to a cloud service.
Source and Relationship Review
Metis also extracts visual embeddings from incoming images and selected video keyframes. These embeddings allow Privacy-Gate to compare visual content even when ordinary file hashes no longer match because the material has been cropped, recompressed, resized, screenshotted or screen-recorded.
Local Multi-Model Workflow
A single incoming clip can require several visual operations:
- face detection for privacy protection;
- licence-plate region detection;
- feature extraction for archive comparison.
Metis accelerates these neural-network workloads, while the host application manages file records, source references, similarity indexing, mask rendering, reviewer controls and approved export.
This division makes the edge architecture clear:
Metis understands what is visible in the media; the host manages how that information is reviewed and acted upon.
The result is:
- privacy protection before cloud upload;
- local processing of sensitive raw media;
- offline-capable review after setup;
- lower host inference load;
- measurable video-processing performance;
- visual archive retrieval without relying on an external service.
Why Voyager Wingman
Voyager Wingman will turn the project requirements into working Voyager SDK pipelines for Metis.
It will be used to generate, run and refine the face-detection, licence-plate and visual-embedding pipelines, then iterate on them as they are tested with real images and short videos.
The public development journey will show how the original prompts evolve into a working multi-model edge-AI demonstration, including the changes required when initial pipeline results do not perform as expected.
Team
Privacy-Gate will be developed by a two-person team :
- Undergraduate Student at the Department of Information and Communication Systems Engineering, University of the Aegean, completing the third year and entering the fourth, responsible for the Metis integration, application development and system demonstration;
- a supporting team member with a graduate Computer Science background, contributing to software design, testing and technical review.
