Redacting an image so it cannot be undone

A black rectangle over a face and a face that has been removed look identical. Only one of them is a redaction.

The three failures

Redaction goes wrong in three ways, and all three produce a file that looks correct.

1. The box is a separate layer

Draw a black rectangle in a PDF editor, a word processor or a design tool and you have added a shape on top of the content. The original is still in the file. Selecting the text underneath, deleting the shape, or opening the document in a different program brings it straight back — and in a PDF, copying the page text often ignores the shape entirely.

This has produced a long series of real disclosures: court filings, government reports, corporate documents. It is not an obscure failure. It is the default outcome of doing the obvious thing.

The only reliable fix is to flatten: produce a new file built from pixel data, with no layers to peel back and no text to select.

2. The pixels are covered but the file still tells you everything

A photograph carries far more than the picture. A typical JPEG from a phone contains the GPS coordinates where it was taken, the make, model and often the serial number of the camera, the time to the second, and frequently an embedded thumbnail of the image as it was before any editing.

That last one defeats the entire exercise. Cover a face, save the file, and a small copy of the uncovered original may still be sitting in the metadata. Some editors update the thumbnail; some do not.

Exporting through a canvas removes all of it, not by scrubbing but as a consequence: the new file is built from pixels, and none of that is pixel data.

3. The area is blurred or pixelated instead of covered

This is the most common choice and the most misunderstood, so it is worth its own section.

Why pixelation is not redaction

Pixelating averages a region into blocks. The crucial property is that this is deterministic: the same input always produces exactly the same output.

So an attacker does not need to reverse anything. If they can guess what kind of thing was written — a name, a card number, a postcode, a date, a word from a limited set — they can pixelate every candidate the same way and compare the results until one matches. It is a lookup, not an attack.

Text is the easiest case of all. The character set is tiny, the font is usually standard, and the layout is regular. Tools that do this exist and are freely available, and they have been used on real published documents and broadcast footage.

Blur is the same story with more mathematics behind it. A light blur is close to reversible. Even a heavy one preserves the length, rhythm and rough shape of what was there — which is often enough, because you rarely need to read a redacted field to identify it, only to narrow it down.

Use blur and pixelation to show that something has been removed. Never to hide what it said.

What a solid block actually guarantees

Filling a region with one colour replaces every pixel in it. There is no averaged version of the original left, no residual structure, nothing to compare against a candidate. The information is gone from the file in the strict sense: it cannot be recovered because it is not there.

That is the only option in this category that makes a promise worth anything.

Checking rather than trusting

Everything above is a description of what should happen. Verifying is different, and it is worth doing because a drawing routine with an off-by-one error produces a redaction that is one pixel short and looks perfect.

The check is simple: decode the exported file again — not the editing canvas, the actual file a recipient would receive — and compare it against the original pixel by pixel. Three questions get answered. Did every pixel inside the covered area change? Is that area now a single flat colour, which only a solid fill achieves? And did anything outside the covered areas change, which would mean the whole picture was re-encoded?

It takes a fraction of a second and turns a claim into a measurement.

What to cover, and how much

Cover generously. A box drawn tightly around a name still reveals its length and rough shape, and on a form the position of a field can identify it as reliably as its contents.

Think about what is left. A redacted street name next to an unredacted house number and a clear photograph of the front door is not redacted. Identification usually comes from the surrounding detail rather than the piece you covered, and this is the failure that catches careful people.

Watch the filename. passport-redacted.png published next to passport.png tells anyone looking exactly what they have found, and the original name frequently names the document.

Prefer a lossless format. A lossy format re-encodes the entire picture, which changes pixels you did not ask to change and leaves visible artefacts around hard-edged blocks.

Frequently asked questions

Can a black box be removed from my image?

Not if the image has been flattened to pixels — there is no layer to remove and no text underneath. It very much can be if the box was drawn as a shape in a PDF or a design file, which is how most real redaction failures happen.

Why is pixelating text unsafe?

Because pixelating is deterministic: the same text always pixelates to the same blocks. Anyone who can guess the sort of thing that was written can pixelate every candidate and compare until one matches. No reversal is needed, and text is the easiest case because the character set is small.

Is a strong blur good enough?

No. A heavy blur still preserves length, rhythm and rough shape, which is often enough to narrow down what was there. Use a solid block for anything that must not be recovered.

Does redacting remove the photo's metadata?

It does when the export is rebuilt from pixels, because nothing that is not pixel data survives that. This matters more than it sounds: a JPEG frequently contains an embedded thumbnail of the picture as it was before editing, which would defeat the whole exercise.

How can I check the redaction actually worked?

Decode the exported file and compare it against the original pixel by pixel — the exported file, not the editing view. Confirm every pixel in the covered area changed, that the area is a single flat colour, and that nothing outside it changed.

Should I use PNG or JPEG?

PNG unless file size is a real problem. JPEG re-encodes the whole picture, altering pixels you did not intend to alter and leaving artefacts around the edges of a hard block.