Cookie consent

This site uses cookies that need consent. Learn more

Skip to content
Insights

How does AI face blurring work?

Overhead view of a busy pedestrian crosswalk with a group of people walking across the street. Faces are blurred for privacy, and one person in the crowd is highlighted with a white circle. Road markings, traffic signs, and long shadows are visible.

How face-blurring software finds and tracks faces in video, what makes a blurred copy safe to hand over, and five checks you can run on any tool with your own footage.


AI face blurring looks for every face in every frame, follows each one as it moves and covers it, and then a person checks the result. That person looks for any face the software missed. Whether the copy is then safe to hand over also depends on the cover: it has to be one that can't be seen through, even by software.

This article explains how those stages work, how they can go wrong, what else in footage can identify someone, and five checks you can run on any tool, using your own footage.

#How does software find a face?

Treating each video frame as a still image, the software marks the areas it judges to be faces – usually by drawing a box around each. This first step is called detection. A detector learns what a face looks like from tens of thousands of example images in which people have drawn a box around every face: one widely used research collection, WIDER FACE, labels 393,703 faces in 32,203 images.

On a new frame, it assigns each box a confidence score, and the tool keeps only boxes that score above a threshold. Set the threshold lower and the tool covers more real faces, but also things that aren't faces; set it higher and it raises fewer false alarms, but misses more faces. In a copy you're handing over, a false alarm costs a blurred patch of wall. A miss discloses a person. So ask the vendor how its detection sensitivity is set, and whether you can change it.

Detection isn't recognition – detection finds the faces in a frame, while recognition matches a face to a particular person, if they are already on record. A blurring tool needs to know where a face is, not who it is. So, if you are assessing a tool, ask the vendor whether it identifies anyone, and whether it keeps anything about the faces it finds once the job is done.

The difference matters in law, too. Under UK GDPR, biometric data becomes special category data, with extra conditions on its use, only when it's used to uniquely identify someone. Finding a face so that it can be covered doesn't do that.

Detection is usually more difficult on CCTV than on phone footage – CCTV cameras are often mounted high, and look down on their subjects. Much of their footage is low-resolution, and the light may be poor. The environment may be a busy station, ward or shop floor, with dozens of faces in the frame, some of them partly hidden. Each of these increases the chance that a face is missed.

So does the size of each face. A face only a few dozen pixels tall gives a detector very little to work with: the researchers who built WIDER FACE found faces 10 to 50 pixels tall far harder to find than larger ones. A camera mounted high to cover a whole concourse spreads its picture across a wide scene, so each face gets few pixels. And a face too small for the software to find may still be large enough for someone who knows the person to recognise. Ask the vendor for the smallest face its tool reliably detects, in pixels, and check the size of the faces in your own footage.

#How does it keep a face blurred when the person moves?

Failing to detect a face is the most serious issue for a tool that blurs faces. If you hand over a redacted video in which a face is missed for a few frames, simply because the person turned their head or walked behind a pillar, those frames could be enough for someone to recognise them.

Detection works one frame at a time and can lose a face when a person turns side-on, is partly hidden or smeared by movement. Tracking closes the gap: it follows each detected face from frame to frame, using where the face was and how it was moving to keep the cover on it in the frames where detection alone would lose it. This is why tracking matters as much as detection. But it can only follow a face that detection has found.

Tracking can fail, too. Two people crossing in the frame, a face leaving the frame and then returning, or a cut in the footage can all fool it. Then the cover drops, or moves to the wrong person. A tracker predicts where each face will be in the next frame and looks for a detection there; some trackers also compare what each person looks like, which helps them hold on to someone through a short gap. 

When two people cross, their predicted positions overlap, and a tracker relying on motion alone can't tell which detection belongs to whom, so the cover can swap from one person to the other. Comparing appearance reduces these swaps – by 45% in one widely cited study – but doesn't remove them. And when detection loses a face for longer than the tracker allows, the track ends and the face is left uncovered.

#Is a blurred face actually anonymised?

It depends on how it was covered, and whether the cover can be removed from the copy that is handed over. There are three common ways to redact a face:

  1. Pixelation replaces it with a grid of squares

  2. Blur smears it

  3. A solid mask leaves none of it

Each works on the pixels inside the face's box. Pixelation divides the area into a grid and fills each square with the average colour of the pixels inside it, so bigger squares preserve less detail. Blur replaces each pixel with a weighted average of its neighbours, softening detail while keeping overall shapes and shading. A solid mask replaces every pixel in the area with one colour, so nothing inside the box survives.

Software can use whatever is left. Back in 2016, in Defeating Image Obfuscation with Deep Learning, McPherson, Shokri and Shmatikov used neural networks to identify faces through pixelation and YouTube's blurring. How did they do that? They didn't reconstruct the faces or remove the blur; their networks recognised who was behind it. So a face can look anonymised to a person watching the processed video, yet software can still identify it.

The European Data Protection Board's Guidelines 3/2019 on processing of personal data through video devices (adopted on 29 January 2020) say what to do. Where other people can be identified in footage released for an access request, 'that part of the material should be anonymised (for example by blurring the copy or parts thereof) before giving the copy to the data subject' (paragraph 97).

On erasure, the guidelines say 'for instance blurring the picture with no retroactive ability to recover the personal data that the picture previously contained, the personal data are considered erased'.

In practice, this means:

  1. The cover needs to be burned into the exported file so that it is part of the image itself. Burned in means the software changes the pixels of every frame, then re-encodes the video, so the original pixels under the cover aren't stored anywhere in the exported file. The cover mustn't be a layer or an effect that can be removed or switched off, leaving personal data recoverable. You should keep the original file and store it separately. Only the redacted copy should be handed over. Our article What is redaction? covers keeping source files while redacting copies, and the risks in a file's metadata.

  2. The kind of cover matters. A blur that software can see through doesn't anonymise anyone. If a tool's documentation doesn't say whether it offers a solid mask, or whether the cover is burned into the exported file, ask the vendor, and test the export yourself.

For UK readers, the ICO's anonymisation guidance explains the difference between anonymisation and pseudonymisation; a disclosure copy is aiming for the first.

#What else can identify someone?

Many things besides a face can identify someone, including tattoos, uniforms, name badges, licence plates and voices – especially to a viewer who knows the people in the footage. Covering them takes more than face detection. Body tracking covers the whole person, not simply the face. Manual blur zones are drawn by hand to cover anything the software doesn't detect automatically, such as badges or screens. Audio redaction prevents people from being identified by their voices. Some tools, such as Identity Cloak, detect vehicles as well as people and blur licence plates as well as faces.

#Why does a person still check the result?

Software can miss something. And once the footage is handed over, it's too late to rework it. A manual review should step through the challenging sections of the footage: typically those containing the things listed above, and anything else particular to your environment. The reviewer should cover anything personally identifiable that the software missed.

In our case study with CIÉ, Paul Whelan, Data Protection Officer at Ireland's state-owned public transport group, says: 'Even in complex scenes, Identity Cloak finds and redacts the vast majority of sensitive material automatically and minimises the need for manual intervention.' At CIÉ, a single video that might have taken four hours to process by hand now takes around 30 minutes.

Reviewing the redacted footage before export is the last of the four steps in our guide to how to blur faces in video for DSAR compliance.

#How can you test a tool on your own footage?

Run these five checks on any tool, including the one you already use:

  1. Use your most challenging footage. Pick a clip where people turn their heads, pass behind pillars or through doors, in poor light or in a crowd. Step through the output frame by frame at those times, and count the frames in which a face can be seen.

  2. Look closely at the cover. View the covered faces at full size. If you can make out the shape of a face, assume software may make out more. Check if solid masks are available.

  3. Open the exported file somewhere else. Open it in a different player or editor to confirm the cover is part of the picture and can't be switched off. Also check that the original is intact and stored separately, and that the exported file's metadata contains no personal data.

  4. Look past the faces. Check the clip for tattoos, birthmarks, scars, badges, uniforms, licence plates, screens, voices and gaits. Then check which of them the tool detects, and whether it lets you redact the rest yourself.

  5. Time the whole job. Time how long the software takes to process the footage and how long a person takes to check and correct its output. Add the two together and compare the total with the time it takes to redact the same clip by hand.

#Where does Identity Cloak fit?

Identity Cloak is Facit's video redaction software. Its AI detects and tracks moving people and vehicles and blurs faces and licence plates. Identity Cloak has no facial recognition. You can raise or lower its blur level, up to a solid mask. The cover is burned into the exported file and can't be removed or reversed, and your original footage is left unchanged. Our guide to face blurring software options discusses Identity Cloak's face and body tracking blurs, one-click manual blur zones, audio redaction and multiple blur types, alongside competing tools.

You can trust automated blurring with a copy you're handing over if the tool passes the five checks on your own footage, and a person reviews the result before export. For Identity Cloak, the paragraph above already answers part of the second and third checks: the cover it offers, and what happens to the exported file and your original. Everything else in the five checks depends on how it handles your footage.

Start your free Identity Cloak trial

If you manage a growing number of footage requests – or are preparing for compliance with the frameworks covered in this guide – Identity Cloak can help you meet your obligations without manual frame-by-frame editing.

Used by police forces, retailers, school districts, NHS trusts and transport operators across the UK and US, Identity Cloak is available as a desktop solution and as a Milestone XProtect plug-in.