Selective redaction: hide the sensitive parts, keep the screenshot readable

There is a tension in every screenshot you have to share. Redact too little and you leak something. Redact too much and you have shipped a picture of black rectangles that proves nothing and helps nobody.

Support tickets are where this bites hardest. The agent needs to see the error, the state of the form, which tab was open, what the sidebar said. What they do not need is the customer’s name, email, account number and the API key sitting in the address bar. Blur the whole upper half and the screenshot no longer answers the question it was attached to.

The two bad equilibria

Blanket blurring. Everything above the fold gets smeared. Safe, useless. This is what people do when they are in a hurry and do not trust themselves to spot every leak, and it is why so many bug reports are functionally empty.

Spot redaction by eye. You cover the four things you noticed. The problem is the fifth — the notification toast that appeared mid-capture, the autocomplete dropdown still open, the account email in the corner of the UI you stopped seeing months ago. Human attention is exactly the wrong tool for exhaustive search over a busy image.

Both failures come from the same place: deciding what to hide is a detection problem, and people are bad at detection over cluttered visual fields.

What targeted detection does instead

The approach in SnapShield AI’s Smart Blur is to find the sensitive regions specifically, and leave the rest of the image untouched and sharp. Two detectors run:

Faces, via an on-device ONNX model. This is the one genuinely machine-learned component in the tool — a face detector that returns bounding boxes, running locally against the image with no network call.

Sensitive data, via the Windows OCR engine plus structural validation. OCR reads the text regions; each candidate is then matched against patterns that verify shape rather than merely looking plausible. Card numbers are checked with the Luhn algorithm. IBANs are checked with mod-97. National IDs are matched by their actual formats — a US SSN’s 000-00-0000, a UK NINO — or by a label next to a value, so Passport No: X1234567 is caught even though X1234567 on its own is just a token.

That validation step matters more than the pattern list. An earlier revision of this detector matched “any 9–12 digit number” as a national ID and flagged order numbers, row counts and timestamps. A detector that fires constantly gets switched off, and a detector that is switched off protects nothing.

The current detector covers email addresses, phone numbers, credit cards, IBANs, bank details, JWTs, API keys — bare, bearer, labelled and vendor-prefixed forms — national IDs, passwords, usernames, IP addresses, physical addresses and personal names.

Crucially, generic visible text is deliberately left alone. The point is not to blur text; it is to blur the text that identifies someone. The error message stays readable.

Blur is the wrong verb for the sensitive case

Worth being blunt about this, because it is the thing most likely to hurt you.

Smart Blur applies a heavy Gaussian to the regions it finds, with a feathered edge so the patch does not look pasted on. That is the right treatment for softening a busy background or a face in a screenshot you are publishing casually.

It is not the right treatment when the underlying data must be unrecoverable. A blur is a convolution with a known kernel, and it has an inverse — the original content can be approximated back out of the output. The mechanics are here, and they apply to our blur exactly as much as anyone else’s.

When the requirement is removal, use solid redaction, which replaces the pixels rather than mixing them. In SnapShield that is the Redact tool and AI Redact’s solid style, and we hold both to a stricter standard than “looks gone”: pixel-level tests fail our build if a redacted region still correlates with what was underneath it. We shipped 1.6.2 as a security release precisely because an earlier build failed that bar.

So: Smart Blur for tidying, solid redaction for protecting. If you would be unhappy seeing that region reconstructed, blur is not what you want.

Detection is a first pass, not a review

No detector finds everything, and the ways it fails are worth knowing before you rely on it.

• Faces at an angle, partly occluded, very small, or in poor light are missed by face detectors generally. This one is no exception.
• OCR is the ceiling on data detection. If the OCR pass cannot read the text — low contrast, unusual font, heavy compression artefacts, a screenshot of a screenshot — the pattern matching never sees it. There is nothing to match against.
• Windows OCR needs a language pack. If the relevant one is not installed, detection quietly finds nothing and reports “nothing sensitive found”. That reads like the image is clean when it actually means the detector never ran properly. If you get that result on an image you know contains an email address, check your Windows language settings before trusting it.
• Context is not detected at all. A licence plate, a name badge, a whiteboard, an internal hostname in the URL bar, a door number — none of these are faces or pattern matches. They are yours to catch.

The honest framing is that detection removes the tedium of the obvious cases so your attention is free for the non-obvious ones. It does not replace looking at the image.

A workflow that holds up

1. Run detection first. Let it mark the faces and the pattern-matched data.
2. Read the image as a stranger. Window titles, browser chrome, toasts, sidebars, counts, reflections, the filename.
3. Choose the treatment per region. Solid redaction for anything that must not be recoverable; blur only where the goal is tidiness.
4. Export flattened, to PNG or JPEG.
5. Verify the export. Open it in a different program, zoom to 400%, push brightness and contrast to the extremes. Solid regions stay flat. Anything that ghosts was not removed.

That last step takes fifteen seconds and is the only one that tells you the truth.

Related reading

• Can blurred text be recovered?
• How to redact a screenshot on Windows
• Redacting screenshots for GDPR