Hiding faces in screenshots and photos: what counts as enough

Blurring a face is the single most common redaction on the internet, and one of the least effective. Not because the blur is weak, but because a face is not the only thing that identifies a person, and blur is not removal.

Two separate failures

The blur is reversible. A Gaussian blur is a convolution with a known kernel, and it has an inverse. Given the output and a guess at the radius, deconvolution reconstructs an approximation of the input. Pixelation is worse: each block is the exact average of the region under it, which turns recovery into a search over candidate faces rather than an approximation. Against a constrained set of possible people — a company’s staff page, a class photo, a small town — that search is short. The general mechanics are here.

Even a perfect removal may not anonymise. This is the part people underestimate. Face recognition is one route to identity among many, and the others survive a black box over the face:

• Build, posture and gait, which is why blurring faces in video is weaker still.
• Hair, tattoos, jewellery, glasses, scars — often more distinctive than facial geometry.
• Clothing, which places a person at a time and event when cross-referenced with other photos.
• Context. A name badge, a desk, a house number, a car, a reflection in a monitor, a street sign.
• Metadata. GPS coordinates in EXIF identify a home address regardless of what you did to the pixels.
• The other people in the frame, who are identifiable and place the blurred person in a known group.
• The surrounding text. A screenshot of a profile with the photo blacked out but the bio intact is not anonymous.

So “did I cover the face” is the wrong question. The right one is “could a motivated person work out who this is from everything left in the file?”

What actually removes a face

Cover it with a solid, fully opaque shape. Not blur, not pixelate, not 90% opacity — that last one leaves a tenth of the original signal, which a levels adjustment brings straight back. A flat fill is many-to-one: every possible face under it produces the same output, so there is nothing to invert.

Cover generously. Hairline to chin, ear to ear. A box tight to the eyes leaves jaw and hairline, which are enough for both algorithms and acquaintances.

Crop instead, when the person is at the edge. Removing the pixels beats replacing them, and there is nothing left to argue about.

Flatten on export. PNG or JPEG, one layer. If the shape stays an object in a project file, anyone can move it aside.

Strip metadata, especially GPS. Then check the filename.

Then read the rest of the image as a stranger would. Badge, screen contents, window titles, whiteboard, the view out of the window, the other faces. Redact those too or crop them out.

When you need the face there but not the identity

Sometimes the picture is about a person being present — a busy office, a conference, a crowd — and a black rectangle ruins it. Options, roughly in order of safety:

1. Do not publish that frame. Pick another shot.
2. Replace the face entirely with a flat shape or a synthetic face. Removal, not obscuring.
3. Shoot differently. Back of head, out of focus at capture, or from a distance where the detail was never recorded. Optical blur at capture time is genuinely destructive in a way that software blur is not.
4. Get consent and skip the problem.

What is not on the list is a heavier blur. A stronger blur is still an invertible operation; you have changed the parameter, not the category.

For screenshots specifically

Screenshots have their own version of this. The face is usually a small avatar next to a name — so the avatar is not what identifies the person, the name is. Blacking out a profile picture and leaving @sarah.chen beside it accomplishes nothing.

Work through the whole surface: display names, handles, email addresses in the account menu, the signed-in user in the corner, notification toasts from third parties, and the list of recent contacts in the sidebar. In a video call screenshot, that means every tile, not just the speaker.

Doing it at volume

One photo is a manual job. A folder of a hundred support screenshots is not, and this is where people start cutting corners — which is where leaks come from.

SnapShield AI exists for that case. Face detection runs on-device with an ONNX model, proposes regions, and you accept or adjust them; batch redaction applies across a folder. Redaction replaces pixels on export, verified by pixel-level tests in our build.

Two honest caveats. Detection is a first pass, not a review — no model finds every face, particularly at an angle, partly occluded, small, or in poor light, and you are responsible for the ones it misses. And detection does not address context: it finds faces, not the name badge, the licence plate or the address bar. Those are still yours to catch.

The check

Export, then open the exported file in a different program. Zoom to 400% on each covered region. Push brightness and contrast hard, and pull levels endpoints together. Flat means it held.

Then do the harder check: hand the file to someone who was not involved and ask them who they think the person is, and how they worked it out. That test catches the context leaks a histogram never will.

Related reading

• Can blurred text be recovered?
• Redacting screenshots for GDPR
• Redact a screenshot without uploading it anywhere