Why blurring sensitive information is a bad idea (2007)

http://dheera.net/projects/blur

by Schiphol • 12 years ago

265 120 12 years ago

dheera.net

Udo4• 12 years ago

A nice Gaussian blur would probably be fine, it's specifically the pixelation technique that's leaking data.

nick_riviera• 12 years ago

Isn't that still forwardly deterministic i.e. you can apply the same algorithm to it but at a higher resolution?

jarman2• 12 years ago

Yes it is reversible for practical purposes. There is some data loss on edges, which will manifest as noise on recovered image.

ekr• 12 years ago

You think the technique mentioned in the article doesn't leak data? It actually leaks more data than a gaussian blur.

Gaussian blur suffers from exactly the same problem, although a different difference function is needed for it.

goldenkey4• 12 years ago

Gaussian blur is equivalent to an auto key cipher [0] without a key. Each pixel is composed of a couple pixels from the original. With some simple algebra, cross-examining pixels that have values determined by the same origin pixel, you can reverse the operation.

[0] http://en.m.wikipedia.org/wiki/Autokey_cipher

Udo4• 12 years ago

I see that I deserved a schooling on gaussian blur, but

  > You think the technique mentioned in the article doesn't leak data?

yes, I do think it leaks data, which was the correct half of my comment, actually.

Not that it matters now, but I always assumed blurring distorted the data in way way that would make it hard to recover in practice.

Qantourisc• 12 years ago

Gaussian blur is a 100% reversible! You just sharpen using the same Gaussian settings for a Gaussian sharpen. It's only due to rounding to 8bit per colour that causes it to loose any information.

jarman2• 12 years ago

It will be worse than that. Pixellation destroys information, it's not reversible in general. But in this case, what is left is enough for recovery, as source domain (check number) is small. Gaussian blur is theoretically lossless (rounding loss and loss on image edges in practice) - you can take arbitrary image, blur it, reverse and get original+noise

Udo4• 12 years ago

Thanks for making me look deeper into what's going on with common blurring algorithms. My instincts were clearly wrong here, for years I thought reversing blurs wasn't feasible in practice. I was very, very wrong.

mxfh• 12 years ago

So if something is 1337 days old it gets autoreposted? Could live with that.

Previous discussion: https://news.ycombinator.com/item?id=1939607

A bit more precise this post is even older, and was first discussed in 2007 prominently in these two places: http://www.reddit.com/comments/xaae/how_to_extract_personal_...

https://www.schneier.com/blog/archives/2007/01/how_to_recove...

(please refrain from responding with XKCD references, I'm aware of that, just want to link to older discussions)

vog• 12 years ago

My favourite comment from the old discussion came from andrewreds:

WHAT... why would you completely black out the number, where you could instead use random coloured squares, that look like it is a blurring, so someone can go through all the effort, decoding your white noise, and thinking in the end they have your number... when they don't ;)

https://news.ycombinator.com/item?id=1939807

CreakyKnuckle• 12 years ago

To be fair, that's a pretty numerically apropos age for HN.

dheera6• 12 years ago

I wrote this a lot more than 1337 days ago. IIRC it first appeared on Slashdot.

http://it.slashdot.org/story/07/01/07/1352242/blurring-image...

Not sure what prompted someone to repost it to HN today. But that's cool because I see people continue doing this on a daily basis.

erikb2• 12 years ago

In some cases even blurring faces might be a bad idea. Just because we are unable to unblur a face today doesn't mean we are unable in 10 years or 100 years. In many cases this might not be a problem but in some cases this might lead to trouble later on which can be avoided just as easily.

goldenkey4• 12 years ago

You cannot unblur mosaic blurring. It's the equivalent of a hash. The best you can do is brute force possible input vectors. There will be many collisions. This technique only works because digits/numbers limit the input space for a credit card or bank number. For faces, the best you can do is validate if someone you already suspect or someone you have in a database, is the origin of the mosaic. If you had a picture of every person in your country, you could run them all through and find the origin - but you're not really unblurring as much as extracting information to do process of elimination on a 'image hash'

erikb2• 12 years ago

Maybe the scientifically correct definition of "to unblur" is reverting the process directly by applying a mathematical algorithm. In a normal conversation like what we have here, bruteforcing a good enough result can also be considered "to unblur" because for the person it's the same result: Everybody knows who it is.

asgard1024• 12 years ago

I can imagine you can computationally unblur mosaic on video, if the object (e.g. face) is not changing, just the camera is moving around a little.

kostasjust• 12 years ago

i agree with erikb.. In theory at least, it seems possible to me that unbluring mosaic could happen if you have in multiple mosaic images from different angles (like mosaic in a video) even without a database with the possible faces).

gus_massa2• 12 years ago

You can't use the brute force method to find any "person" you have in the database. You can only find any "photograph" that you have in the database.

If you have a photograph of me, and I just let my hair and beard grow for a month, then you won't match the mosaic version of my hairy version with the old photograph.

Perhaps you can add some filters to adapt the contrast, luminosity, hue, illumination, rotation, pixel shifts and noise. It would be more difficult to fix the head orientation, open/close mouth/eyes.

Robin_Message• 12 years ago

If you have trial faces at the right angle, you can unblur faces, it's just a bit harder to generate. But nothing major is going to change in 10 or 100 years; it's just information theory as to whether you can identify the person/number or not.

bithush2• 12 years ago

Indeed. See http://en.wikipedia.org/wiki/Christopher_Paul_Neil

seanflyon2• 12 years ago

Swirl is not a lossy process. If he had blurred his face he would have been much harder to catch.

nnnnni• 12 years ago

Their answer is to "color over it", but BE CAREFUL if you do that. I'm sure that some of us remember the US government document(s) in the late 90s/early 2000s that were released in redacted form, but the person didn't realize/understand that people with the Acrobat editor program could remove the black bars.

currysausage• 12 years ago

Also great: HTML pages with sensitive passages using #000 for background and foreground. You just need to select and copy to Notepad - voila.

I think I once saw that on the site of a home security company, citing an email from customers who thanked them for burglarproofing their home in XX XXXXXXXX Street. Quite ironic.

MichaelApproved• 12 years ago

You don't have to go all the way back to the early 2000s to find poorly redacted federal documents. Rachel Maddow had a great segment on botched redactions, like removing the black lines from a PDF.

The best solution was to actually cut out the parts of the page you don't want seen and then scan in the result. All other methods are prone to mistakes.

Edit: corrected link to segment http://www.msnbc.com/rachel-maddow-show/watch/redaction-rule...

Houshalter4• 12 years ago

Would a screenshot work?

dheera6• 12 years ago

Along the same lines, I remember an MIT Mystery Hunt puzzle once where there were hidden clues in a PostScript document. If you didn't think to look at the PostScript source you probably had no hope of solving it.

shinigami• 12 years ago

Also, image formats like JPEG have a thumbnail stored in them, which may not be updated when you edit the image.

robin_reala2• 12 years ago

Do they? First I’ve heard of that…

dheera6• 12 years ago

Also, when blacking out data from JPEGs, leave generous margins, because there may be compression artifacts in the nearby pixels that also act effectively as a one-way hash. I haven't tested how consistent these artifacts are, but it's better to be safe than sorry.

danielweber• 12 years ago

There is also "Leaking Sensitive Information in Complex Document Files--and How to Prevent It" from IEEE Security & Privacy earlier this year. All copies might be behind paywalls, however.

matchu• 12 years ago

For a fun variant of this problem, see the 2008 Underhanded C Contest: http://underhanded.xcott.com/?page_id=17

bithush2• 12 years ago

I feel this is also very relevant http://en.wikipedia.org/wiki/Christopher_Paul_Neil

Police took a photo with a "swirl" effect of a paedophiles face and reversed it to reveal a very usable picture. So good in fact he was found and arrested.

danepowell• 12 years ago

He wasn't just a pedophile, he was a child molester.

frogpelt• 12 years ago

"just a pedophile"

A phrase I never thought I'd see.

jebus9892• 12 years ago

Police == 4chan?

exelius• 12 years ago

Yeah, basically. 4chan's admins have been working with the police for many years because of the questionable shit that gets posted on /b/ (child porn, murder threats/evidence, suicidal posts, etc.) They would have been shut down years ago if they weren't cooperating.

topdownjimmy• 12 years ago

Also relevant?

https://www.youtube.com/watch?v=Z3xd8bXlzeU

https://www.youtube.com/watch?v=xYubZoycACw

nathell• 12 years ago

In some cases, coloring over parts of the image might still not be enough. Specifically, when all of the following are true: (i) the domain of possible entries is reasonably small (e.g., a number or a name and surname), (ii) the text is printed in a proportional (not fixed-width) typeface, and (iii) enough of the rest of the line is visible to infer the font size and kerning settings.

userbinator• 12 years ago

Actually, using fixed-width would be even easier since it reveals the length.

bryondowd• 12 years ago

I'd be curious to see some numbers backing one or the other, but I suspect proportional would be better. Fixed width narrows to a number of characters. So it depends how many possible values have x characters. Proportional narrows to a particular width, which could be more granular than character count. So it depends how many possible values turn out to be exactly x pixels across given the type settings.

ColinDabritz• 12 years ago

I think this form of analysis is amazing. It breaks the "The best solution was to actually cut out the parts of the page you don't want seen and then scan in the result." approach from elsewhere in this thread as well. Even cutting out each portion of each line isn't enough because you can reconstruct line sizes. You could cut out each individual word and string them together, with a single size "redaction" block perhaps, but that's a lot of work.

tetrep• 12 years ago

You could cut out each individual word and string them together, with a single size "redaction" block perhaps, but that's a lot of work.

Not really. OCR is plenty good enough to scan a document and then you can just replace anything you wanted redacted with a common string: [REDACTED]

Even if OCR wasn't an option, it would still (probably) be easier to type up the document than it would be to physically cut out all the things you want to redact.

hawleyal• 12 years ago

Not really. You would need to have a copy of an uncovered version to know how many lines of text, font size, kerning et al.

DanBC• 12 years ago

Some documents are trivially easy to get near copies.

Other documents provide their own context - they have a lot of other text - that you can use to get font, font size and spacing and etc.

bjourne• 12 years ago

Has this attack actually been proved possible? He writes that he thinks it should work but doesn't have the time or inclination to prove it. If anyone wants to take a stab at it, I'll gladly submit a mosaiced photo of my credit card because I don't think the attack is practical. If you crack it, you're free to keep whatever you can get. :)

bluedino2• 12 years ago

You have to know the original font the numbers were in, and the original algorithm or program used to do the blurring. Then it's just a matter of brute forcing it.

Using a 'smear' brush would be a more secure way to blur, there are various sizes and you introduce some randomness in the path of your mouse/stylus when blurring, and you can go over the numbers multiple times. Blurring to much higher factors is a better idea but it is simpler to just use black bars.

dheera6• 12 years ago

You have to be careful doing this. If your blur tool happens to be of the variety that conserves total brightness (i.e. the blurring is accomplished by 2D convolution with a kernel that has a volume of 1, so sum of all pixel values remains constant before and after blurring), it can still be dictionary attacked pretty easily no matter which random ways you blur, since each digit in the cheque font probably has a different amount of total black area, and your blurring preserves that information perfectly.

If you blur tool is of the variety that doesn't conserve total brightness, such as most "smudge" tools, and you use human randomness to blur it, then it would probably be pretty hard to reverse.

I still recommend cutting out the sensitive information rather than blurring it, just to be safe. Also, leave a generous margin, (1) to avoid giving information about length, and (2) because lossy compressors may have left tiny artifacts of the sensitive information in the areas around it.

cbhl• 12 years ago

In North America, all the checks (cheques) use the same fonts. See http://en.wikipedia.org/wiki/Magnetic_ink_character_recognit...

IvyMike• 12 years ago

Better yet: change the text to "you think you're clever, don't you" and then blur the image.

It's a little easter egg for the people with an unblur plugin.

Houshalter4• 12 years ago

Web archive version: https://web.archive.org/web/20140714183916/http://dheera.net...

wcummings• 12 years ago

I entered an underhanded contest one year where the challenge was recoverable-but-correct-looking redaction of jpegs. I used an insecure random seed based on the time (you could brute force an unredacted image based on the rough time it was generated). The winning solutions were more clever

dalke2• 12 years ago

You (and I) were among the many which implemented that exact technique for http://underhanded.xcott.com/?page_id=17 . :)

cookiecaper• 12 years ago

Most people want to blur because they think it makes their photo flow better than gaudy black highlights, but you can also use highlights that match the background color/image and make it blend it in, leaving a big white space on a white background instead of a big black space. Most people won't notice anything is missing at all.

If you want that pixel mosaic look for extra futuristic feel, remove the original content by making it fade into background, replace with new, irrelevant content, and then pixelize.

tshadwell• 12 years ago

And when you colour in the picture always use a pen of 100% opacity or the colour can be removed from the image to reveal the data underneath!

georgemcbay• 12 years ago

I think a lot of people are unaware of how easily you can achieve blind deconvolution on many images blurred with most blur algorithms and even real-world blurring effects (including motion, out-of-focus, etc blurs).

The results won't be perfect, but they are usually close enough to see much of the detail that appeared to be lost.

I never use blur to obscure sensitive information; black that shit out (and then also make sure you aren't saving it as metadata or in a layer) or just replace it with fake data.

yeukhon• 12 years ago

Wouldn't cutting that hidden text out entirely with a photo editing tool be better? I've heard people talking about recovering text f. Well, in the case of edges sticking out of the black bar (this happens a lot for people who aren't careful) like using paint, I believe there is a chance to recover partial, if not, the entire content back.

Also, doesn't adding a black bar on top of a text means just adding more bytes to the file, instead of removing the bytes belonging to the now hidden text?

ohazi• 12 years ago

I'll just leave this here:

http://refocus-it.sourceforge.net/

JoshTheGeek• 12 years ago

http://blog.mailgun.com/open-sourcing-our-email-signature-pa... (https://news.ycombinator.com/item?id=8081532) has some screenshots with blurred email addresses you can read without fancy deciphering.

jheriko• 12 years ago

do those pixel blocks actually represent accumulated samples from the image?

any number of very obvious methods can be used to avoid this besides using a black box...

cnst• 12 years ago

Where did they find such a crappy font for body? SimHei for Latin letters looks absolutely horrendous with disabled font blurring.

They advise you not to blur, yet require blurring for comfortable reading? How ironic!

Why blurring sensitive information is a bad idea (2007)

Related Stories