Pixelation looks like it destroys information. The text turns into a grid of colored squares you can't read. But looking unreadable and being unrecoverable are two different things. For a large class of real-world data, pixelation is reversible, and so is blur. Here's why, how the attacks work, and what to use instead.
The short answer
Yes, pixelation is frequently reversible, and blur even more so. Neither one removes the information. They both apply a predictable mathematical transform to the original pixels. If an attacker can guess the small set of things the original might have been, they can reproduce that transform on each guess and find the one that matches. That's the whole attack.
The core problem: pixelation and blur are lossy but deterministic. The same input always produces the same output, so the output leaks which inputs are possible.
Why pixelation is reversible
Pixelation (mosaic) works by splitting a region into blocks and replacing each block with the average color of the pixels inside it. That average is a specific, repeatable function of the original content. Now consider redacted text that has a limited, structured format, such as a credit-card number, a date, a license plate, an amount or a national ID. An attacker can do the following:
- Take the known font, size and position of the redacted text.
- Render every plausible value, meaning every digit string or plate combination.
- Apply the same pixelation grid to each candidate.
- Compare each pixelated candidate to the redacted region and keep the closest match.
Because the block averages differ between, say, a "3" and an "8", the candidate that reproduces the exact block pattern is almost certainly the original. The open-source tool Depix famously demonstrated this on pixelated mono-spaced screenshots, recovering readable text from the mosaic alone.
Why blur is no safer
A Gaussian blur is a convolution, which is a filter that can be inverted mathematically (deconvolution). In perfect conditions it's fully reversible. In practice, noise and clipping make exact inversion hard, but they don't save you when the content is predictable. Researchers have shown that blurred or pixelated faces and text can be re-identified by training a model to map obscured images back to their originals, or simply by brute-forcing a known value space. For anything short and structured, blur narrows the answer to a handful of candidates, and sometimes to exactly one.
When is pixelation "probably fine"?
Reversibility depends on how much the attacker can guess. Heavily pixelating a long, high-entropy secret with a coarse grid leaves little to reconstruct. But you rarely know in advance how predictable your data is, and the failure mode is silent: the image looks redacted whether or not it actually is. That's exactly the kind of risk you don't want to carry on a card number, an address or an ID. The safe default is to not rely on obfuscation at all.
What actually works: solid boxes
The only redaction that can't be reversed is one that removes the pixels rather than transforming them. A solid, fully opaque fill overwrites the region with a single flat color, so there's no residual signal correlated with the original and nothing to reconstruct. Then flatten the image on export so the box isn't a movable layer. This is why Privshot defaults to solid boxes rather than blur or mosaic: the output carries zero information about what was underneath.
We go deeper on the general principle in why blurring and pixelating are not safe redaction.
A quick rule of thumb
- Solid opaque box plus flattened export: irreversible. Use this.
- Blur or pixelate: treat as reversible for any short or structured data.
- Movable black rectangle in a PDF or design file: not redaction at all until flattened.
If you're redacting screenshots regularly, see the practical walkthrough in how to redact a screenshot the safe way, or automate it with the Privshot API.
Frequently asked questions
Is pixelation reversible?
Often yes. Pixelation is a deterministic average of the original pixels, so for short, predictable text like card numbers, dates or license plates, an attacker can search for the input that produces the same pixelated output and recover it. Tools such as Depix demonstrate this on mono-spaced text.
Can blurred text be recovered?
Blur is a reversible mathematical operation in principle, and in practice blurred text with a limited set of possible values can frequently be reconstructed or narrowed to a few candidates. It shouldn't be trusted for sensitive data.
What is the only safe way to redact an image?
Cover the sensitive area with a solid, fully opaque fill and flatten the image on export. This overwrites the original pixels so there's nothing left to reconstruct, unlike blur or pixelation, which merely transform the pixels.
Why can pixelated license plates and card numbers be read?
Because the set of possible values is small and structured. An attacker can pixelate every candidate with the same block grid and match the result against the redacted image, recovering the original by elimination.
In short
Pixelation and blur don't delete information. They scramble it in a repeatable way, and repeatable scrambling of predictable data can be undone. For anything sensitive, cover it with a solid, opaque box and flatten the export. You can redact a screenshot safely here.