applets.

Pixelated Isn't Hidden

← All apps

About Pixelated Isn't Hidden

People blur or pixelate a password, a PIN or a name in a screenshot and think it is gone. It often is not. Pixelation replaces each little square of the picture with the average colour of that square: a fixed, repeatable recipe. Because it is repeatable, an attacker can pixelate every possible secret the same way and see which one gives the same blocks. This page does exactly that, only ever on text you type here, drawn by the page itself. Nothing you type is uploaded or kept, and nothing leaves the tab.

RECOVER A PIN: type a four-digit PIN and the page renders it, pixelates it, then pixelates all 10,000 PINs from 0000 to 9999 in the same font and block grid and scores each against your pixelated blocks. The closest match is usually your PIN, found in a fraction of a second, even with blocks wider than a whole digit: with the font, size and grid known it keeps working up to blocks of about 40 pixels, and only past that does it start to fail. Take the block slider to the end to see it lose. Tick attacker must guess the grid to shift the block grid by a secret amount: assuming it lines up usually fails (the page shows what that guess gives), so it tries every offset, and the time line counts every try. Tick add JPEG-style noise to rough up the blocks and watch the winning margin shrink. Attacker's font: the PIN is always drawn in a monospace font. Assume the wrong font (Inter) and every guess leaves different blocks, so the attack fails at every block size; try a short list of three likely fonts and the right one fits far better than the others, so it is found again. When the pixelated text sits in a document whose other text is readable, the font is usually there to copy.

VIDEO: a pixelated patch over a moving picture, the way a screen recording or a news clip blurs something out. The grid stays still while the picture under it shifts a pixel or two each frame, so each frame averages a different set of pixels. The page treats every frame's blocks as equations about the hidden picture and solves them together, a little at a time (the Kaczmarz method), and measures how much of the pixelation that undoes. A sideways pan brings back detail across but not up and down; a shaky hand brings back both. More frames than there are pixels in a block, at different offsets, can undo nearly all of it.

PASTE & DECODE: a fair test of the PIN attack. Make a pixelated PIN, copy the picture (or save it), and paste or drop it into the decoder. The decoder gets only the pixels. It finds the grid by trying every block size and every start across and keeping the biggest size whose cells are flat inside (a pixelated picture is flat inside each block), then scores all 10,000 PINs against those blocks. Nothing leaves the tab.

WORDS: the same trick for a word, too many combinations to try them all. The page draws every letter a to z in each slot of your word, in the same font, pixelates each the same way, and then goes slot by slot, keeping the letter whose blocks, added to its neighbours', come nearest yours, sweeping until nothing changes. The strip shows the alphabet pixelated the same way. Tick attacker has a word list and it also tries every real word of that length from the list Word Solver uses (about ten thousand six-letter words, in a few milliseconds), adding up each word's letter blocks the same way: a real word is found at block sizes where letter by letter has long failed, while a made-up string isn't in the list at all. Depix does the job with one reference image holding every pair of letters (a De Bruijn sequence) and needs exact matches, so it fails if the font, size, colours or grid differ or the picture was re-saved as a JPEG; later tools search the grid and accept near matches, so those are not reliable protection.

BLUR & SWIRL: a blur is an average too, and a known blur can be partly undone by working the averaging backwards (Van Cittert deconvolution); noise sets how far you get. For text that hardly matters: a blurred word falls to the same guess-and-check as a pixelated one. A swirl just rotates the picture more the nearer the centre, so it comes back almost exactly once you know the centre and strength.

HOW TO REDACT: what works. Delete the text itself, then cover its place with an opaque box of any colour, a little wider than the text so its length does not show, and flatten the result (or use a proper redaction tool, which does the deleting). The page simulates the difference between a box painted over live text in a document and one flattened in.

THE STORY

In December 2020 a security consultant, Sipke Mellema, released Depix, a tool that recovers text pixelated with a plain box filter by matching it against a reference image built from a De Bruijn sequence of characters. It drew about 10,000 stars on GitHub within days. In 2016 researchers at the University of California, San Diego (Hill, Zhou, Saul and Shacham) had already shown that both pixelated and blurred text can be read back, even with unknown grid offsets and some noise. In February 2022 Dan Petro at the firm Bishop Fox answered a challenge from Jumpsec with Unredacter, which searches for the grid, and a write-up titled "Never Use Text Pixelation To Redact Sensitive Information". In April 2025 Jeff Geerling pixelated a folder list in a video and offered a reward to anyone who could read it; within a day three people had, because a moving pixelated window leaks a little more of the picture in every frame.

Other hiding tricks have been undone in the field too. In 2007 specialists at Germany's federal police (the BKA) reversed swirled photographs of a wanted man's face. On 8 October 2007 Interpol published the reconstructed pictures in its first worldwide public appeal, and he was arrested in Thailand on 19 October. Redaction fails the other way as well: in January 2019 a court filing by Paul Manafort's lawyers carried black boxes laid over live text in the PDF, and reporters read the hidden words simply by copying and pasting them out, because the text under the boxes had never been removed.

HOW IT WORKS

The page renders your text to a hidden canvas as a map of ink, where 0 is blank paper and 255 is full ink, rounded to whole numbers as a screenshot would be. Pixelation cuts that map into square blocks and replaces each block with its average. To recover a PIN it renders each digit on its own in each of the four slots, takes the block averages, and adds them up for every one of the 10,000 combinations (averaging is linear, so a candidate's blocks are just the sum of its digits' blocks: no re-rendering needed). The combination whose blocks are nearest yours, by sum of squared differences, wins. Averaging is many-to-one, so it cannot simply be run backwards; guess-and-check works because the blocks that remain are still enough to tell the 10,000 candidates apart. A shifted grid breaks the match until the attacker tries every offset, which the page does when you ask it to (and counts every try). Noise only loosens the match. A different font or size would need the attacker to find those too, which real tools also search for.

The swirl is a rotation whose angle grows toward the centre; applying it with the opposite sign undoes it, apart from what rounding to whole pixels loses. The page's picture is tall enough to hold the whole swirl; in a real photo, parts twisted off the edge are gone, and the undo can only repeat the edge pixels there. The slider gives the twist at the centre in degrees. The scores for both undos are measured on the letters only, as the share of the damage put right. The blur undo repeatedly asks "what picture, blurred, would give this?" and nudges toward it; with a box blur, whose averaging flips sign in the frequencies, that nudge runs away, so the page undoes a Gaussian blur and says plainly that it cannot simply reverse a heavy one. Everything here is computed for real on your text and your text alone.

FURTHER WATCHING AND READING

4 digits, made up
Block
Attacker's font
The PIN
Pixelated (hidden?)
Read back