Are you human?

So I've been working on this captcha algorithm recently, and I have a pretty good feeling about it. Unfortunately, I haven't yet had the time to set up a proper testing environment (eg: OCR analyzers and such), but my gut feeling is that it should be sufficiently secure against attacks.

Below, I've attached two sample output files (the input file used a black font over a white background). Notice that in the first one (using the default settings) the smaller fonts are fairly hard to read. In that case, the parameters can be tweaked to sharpen the image (the second snapshot). The image can also be blurred, if necessary.

Anyway, what do you guys think? Will it work?

.....

I think CAPTCHAs are flawed by nature anyway, due to people "beating them" by putting the CAPTCHA on another website (i.e. a porn website) so a real human fills it out.

But they do look good, although they're a bit hard to read -- even for my 20 year old eyes ;). The first is a lot harder to read than the second, it would be good to see how OCR does on the first vs the second. If I didn't know what the words were, I doubt being able to read a few characters would help me. For example "destruction" could easily read "description" IMO.

Can't really read the first image except for the bigger letters. And even then badly. That would be a terrible choice.

The second one... well I can't read the two first lines. The rest I could read. But it would depend on whether you would use normal English or jumbled words. If the latter, probably not. With that font I would probably be confused with the "h" looking like an "n", etc.

I read the second no problem, the first however is pretty tough. Specifically the sixth word on the first line. That word is even kind of difficult to make out in the second one and I might not have gotten it out of context.

Yeah, the value of the parameters definitely depend on the font type and size, as well as the foreground/background colors being used. Choosing the best fit is something of a black art, I suppose.

The three images attached below (pasted together into one) were generated with the same settings (0.2, 0.025), and IMO pretty easy to read (it may help for some to either lean back a bit, or better yet, look at them at a sharp angle). What I'd really like to know is if they really could resist an advanced OCR engine, or if I'm, in fact, seriously underestimating the problem? Well, that's the next step, anyway, and I plan on running it through as many OCR's as possible, for good measure.

...

Quote:

---

Originally Posted by Sebastiani

Are you human?

---

I have not had the opportunity to pass the test of humanity involving the gom jabbar.

Quote:

---

Originally Posted by laserlight

I have not had the opportunity to pass the test of humanity involving the gom jabbar.

---

Hmm, that would've come in handy on my last date (don't ask)... :D

Quote:

---

Originally Posted by laserlight

I have not had the opportunity to pass the test of humanity involving the gom jabbar.

---

I kept my hand in the box.

Fairly unreadable, the first one. However, both of them become a fair share more readable if you open it in gimp and run the despeckle filter with the following settings:
Adaptive: yes
Recursive: no
Radius: 1
Black level: -1
White level: 103

But of course that would also mean that it becomes a fair share more readable to computers by simply running that filter through it. So, all in all, a computer program would have little difficulty reading it, I believe.

Quote:

---

Originally Posted by Sebastiani

Yeah, the value of the parameters definitely depend on the font type and size, as well as the foreground/background colors being used. Choosing the best fit is something of a black art, I suppose.

The three images attached below (pasted together into one) were generated with the same settings (0.2, 0.025), and IMO pretty easy to read (it may help for some to either lean back a bit, or better yet, look at them at a sharp angle). What I'd really like to know is if they really could resist an advanced OCR engine, or if I'm, in fact, seriously underestimating the problem? Well, that's the next step, anyway, and I plan on running it through as many OCR's as possible, for good measure.

...

---

You want a good reliable way to beat the machines? Take a simple one word/character string input, print it to the screen using your noise-a-fier twice along with a third word/character string of equal length (all in random order, of course). Tell your user to type out the word that appears twice.

This not only gives your human user two chances to distinguish the word, but also confuses the machine a bit more by adding a layer of logic instead of just edge detection.

Quote:

---

Originally Posted by Sebastiani

What I'd really like to know is if they really could resist an advanced OCR engine, or if I'm, in fact, seriously underestimating the problem? Well, that's the next step, anyway, and I plan on running it through as many OCR's as possible, for good measure.

---

I'm no OCR expert by any means, but to me the easiest way to read this with OCR would be to acknowledge the simple fact that the readable text is black text on a white background. Running this through a filter that replaces all exposed, non-black pixels with white pixels, and recolors all non-black pixels surrounded by black pixels (all adjacent, or perhaps a minimum 7 of 9 adjacent pixels) to black would yield a very readable image that even the most simple OCR algorithms can read.

Picking the same font face and size makes possible attacks a lot easier. If people find the right noise filter, the result is dead easy for any OCR to deal with.

Make each letter
- a different size (within some limit)
- a different font
- a different face (bold, italic)
- a different orientation

Where letters end up overlapping, experiment with AND/OR/XOR logic for combining pixels.

Quote:

---

Originally Posted by EVOEx

But of course that would also mean that it becomes a fair share more readable to computers by simply running that filter through it. So, all in all, a computer program would have little difficulty reading it, I believe.

---

Quote:

---

Originally Posted by Salem

Picking the same font face and size makes possible attacks a lot easier.

---

Yeah, I thot most of what makes an effective CAPTCHA is not obscuring via filter but warping and overlapping. A lot of them don't obscure at all, they are just very warped.

I agree the first one is harder to read, I wouldn't use something that hard as a CAPTCHA.

I don't think the second one is good though. Think about it, hunt down all the black(ish) pixels, heed only those, trace the edges, and compare the general shape with a database of that single, simple, font that your using.

I would advise using a variety of fonts and colors, and do a lot of warping, resizing, combine it with a similar simple background to further hinder tracing methods. Static like that applied over the whole thing only hurts a machine as much as a human, i.e., don't do it.

Quote:

---

Originally Posted by MK27

Yeah, I thot most of what makes an effective CAPTCHA is not obscuring via filter but warping and overlapping. A lot of them don't obscure at all, they are just very warped.

---

Sometimes too warped. Bs and 8s, Is and 1s, 5s and Ss... they all get very confusing even to the human eye, which is what I feel is the most innate problem with CAPTCHA technology in the first place. In many cases, they demand multiple tries from the end-user and generally offer solutions to continuously bypass complicated CAPTCHAs in order to find a suitable one that you can fill out. The same methods are frequently applied by the tools designed to crack CAPTCHAs.

By the way, I'm curious to know if any one has developed a CAPTCHA cracking tool that uses sound analysis to take advantage of the blind-accessible CAPTCHA forms. Has anyone seen something like this?

Quote:

---

Originally Posted by SlyMaelstrom

I'm no OCR expert by any means, but to me the easiest way to read this with OCR would be to acknowledge the simple fact that the readable text is black text on a white background.

---

Yeah, I was thinking about this too.

For illustration purposes, I ran the second picture in lab mode, extracted the lightness channel and applied a levels adjustment to it (reducing darkness and increasing the midtones). The image below was the result. Fairly easy to OCR. Now, my approach may not even be the easiest. But I think that by using lab mode, I guarantee I can tackle for any kind of noise in the image. As far as coding is concerned, I suppose this could be fairly easily achieved by someone bent on defeating this captcha.

Attachment 9804

Quote:

---

Originally Posted by SlyMaelstrom

By the way, I'm curious to know if any one has developed a CAPTCHA cracking tool that uses sound analysis to take advantage of the blind-accessible CAPTCHA forms. Has anyone seen something like this?

---

I've wondered about this too. I haven't tried, but I think you could refine Sphinx to do that pretty well. For the most part CAPTCHAs hold back individuals, I'm sure groups, like professional spammers, aren't hindered.

I was just thinking about this. Why not use a large set of photographs or drawings of simple objects and ask the user to identify them? I suppose it would be hard to find enough objects that have unique names, but you could couple that with a human language question, such as "Is this a man or a woman?" for one pic, "What color is the bus?" for another, or "How many apples are in the bowl?" for another.

You could generate large sets of such picture/question combinations, and then automate simple changes in the images to prevent someone from simply building a library of correct answers. Just making such a library work would require a lot more complex work than cracking a captcha, and a sufficiently clever system of this sort might be unbreakable.

Quote:

---

Originally Posted by MK27

Why not use a large set of photographs or drawings of simple objects and ask the user to identify them?

---

It is certainly possible, and such initiatives already exist, e.g., ASIRRA.

Quote:

---

Originally Posted by MK27

I was just thinking about this. Why not use a large set of photographs or drawings of simple objects and ask the user to identify them? I suppose it would be hard to find enough objects that have unique names, but you could couple that with a human language question, such as "Is this a man or a woman?" for one pic, "What color is the bus?" for another, or "How many apples are in the bowl?" for another.

---

It's been done and I've been on the user end of it. I can't recall exactly where I've seen it, but I have. The problem with this method is that the data sample would be small enough that one could simply brute force it through prior knowledge of one or more of the questions being asked. Even if you were to have 1,000 unique questions and gave users two attempts before blocking the IP, you'd still only need a relatively modest list of anonymous proxies in order to break it while only knowing the solution to a single question. You could probably even apply the birthday paradox to this one as one might expect at least a few solutions to be a low number between 1 and 9.

Quote:

---

Originally Posted by zacs7

I think CAPTCHAs are flawed by nature anyway, due to people "beating them" by putting the CAPTCHA on another website (i.e. a porn website) so a real human fills it out.

---

How about reverting it? We put porn pictures on the site instead of a captcha and if the user clicks through in less then ($porn) seconds, its a bot and gets kicked :)

Quote:

---

Originally Posted by SlyMaelstrom

The problem with this method is that the data sample would be small enough that one could simply brute force it through prior knowledge of one or more of the questions being asked.

---

Yes, this is a challenge. From the ASIRRA website:

Quote:

---

Past projects have used photographs to tell computers and humans apart. Examples include Carnegie Mellon's PIX CAPTCHA, Oli Warner's KittenAuth, and work done by Chew and Tygar. These projects have a common weakness: they use relatively small image databases. There's a fundamental reason for this. It's difficult for a computer to automatically classify pictures with high accuracy — that's why the task is useful as a HIP. An image database small enough to be constructed manually by a researcher is also small enough to be manually reconstructed by an attacker.

Asirra is different because of our unique partnership with Petfinder.com, the world's largest site devoted to finding homes for homeless pets. They've provided us with over three million images of cats and dogs, manually classified by people at thousands of animal shelters across the United States.

---

Quote:

---

Originally Posted by SlyMaelstrom

It's been done and I've been on the user end of it. I can't recall exactly where I've seen it, but I have. The problem with this method is that the data sample would be small enough that one could simply brute force it through prior knowledge of one or more of the questions being asked

---

Sure, a naive implementation will have this problem. But if you literally did something like generated a large random list of objects regularly (say once a week) then retrieved images with google images (to automate), and let the user have three guesses (or move on to another object) "brute force thru prior knowledge" (this is what I meant by a library of answers) will be a failing venture.

Incorporating the question adds some problems because if the element is simple enough to automate changes to (color or number) then it would also be easy to brute force *unless* you are only given one chance. If the question is truly simple ("How many apples in the bowl?") a human being will get it right the first time, but a bot never will. Of course, you really do have to vary the number of apples in the bowl. I think it will be much easier to present images with numbers of things that will be much more difficult to machine analyze than letters in a CAPTCHA, but also very easy to keep varied (3 pieces of fruit, 4 pieces of fruit, 5 pieces of fruit -- a bowl with an apple, an orange, a banana, and a truck in it, hmmm.)

Vis the birthday problem, that is a problem, but I only spent <5 minutes thinking about it so far :p It would not be hard to start collecting the "really dumb == bot" IP's in that case. If a human being cannot tell a piece of fruit from a truck 90% of the time or better, they aren't using the web anyway and hence will not get unfairly blacklisted.

The ASIRRA solution (asking you to correctly identify a set of pictures rather than just one) looks very good because a human will always get 10 out of 10, they have a constant supply of millions of appropriate photographs (could never be collected for the bot database) and this very much reduces the birthday problem (there are 1024 possibilities just using "dog or cat"). You will have to wait for the year after next for your party ;)

I'd bet black and white shots of dissimiliar objects taken from any and all angles could give you a pretty good database of this sort and be well beyond the potential of most cracking systems -- image analysis on that level is just way too processor intensive. You might as well hire someone at $2/hour somewhere to do it for you at that point.

Quote:

---

Originally Posted by Mario F.

Yeah, I was thinking about this too.

For illustration purposes, I ran the second picture in lab mode, extracted the lightness channel and applied a levels adjustment to it (reducing darkness and increasing the midtones). The image below was the result. Fairly easy to OCR. Now, my approach may not even be the easiest. But I think that by using lab mode, I guarantee I can tackle for any kind of noise in the image. As far as coding is concerned, I suppose this could be fairly easily achieved by someone bent on defeating this captcha.

Attachment 9804

---

Ouch, yeah, even the crudest path-finding AI could connect the dots there quite easily. Unfortunately, when I posted that I had been programming for several days straight, and by the time I got around to creating the sample images, I hadn't put much time into selecting the settings! Can you confirm similar results with the "ibrkfy" sample I posted?

Quote:

---

Originally Posted by Yarin

I would advise using a variety of fonts and colors, and do a lot of warping, resizing, combine it with a similar simple background to further hinder tracing methods. Static like that applied over the whole thing only hurts a machine as much as a human, i.e., don't do it.

---

Quote:

---

Originally Posted by Salem

Picking the same font face and size makes possible attacks a lot easier. If people find the right noise filter, the result is dead easy for any OCR to deal with.

Make each letter
- a different size (within some limit)
- a different font
- a different face (bold, italic)
- a different orientation

Where letters end up overlapping, experiment with AND/OR/XOR logic for combining pixels.

---

Good ideas, thanks!

Captchas are pretty cool. Thats a long sentence to put in though. I like websites that ask you do answer a question to do something like "What's 1 plus four?" or, "How many hands does the average human have?". I wonder if those will gain in popularity in the future.

Quote:

---

Originally Posted by Sebastiani

Can you confirm similar results with the "ibrkfy" sample I posted?

---

I can't. The ibrkfy sample offers a lot more protection against that type of attack, as you can see below. You offer a lot less contrast and you have noise with the same luminance levels as the letters, which doesn't allow me to get rid of it while maintaining legibility for the OCR. I think it can still be read. But you'd need a very good OCR.

Problem is that the "ibrkfy" sample is also a lot harder to be read by a human :(
Especially if you apply some of the good advise above, like warping and different fonts between letters.

Attachment 9805

Quote:

---

Originally Posted by Mario F.

I can't. The ibrkfy sample offers a lot more protection against that type of attack, as you can see below. You offer a lot less contrast and you have noise with the same luminance levels as the letters, which doesn't allow me to get rid of it while maintaining legibility for the OCR. I think it can still be read. But you'd need a very good OCR.

Problem is that the "ibrkfy" sample is also a lot harder to be read by a human :(
Especially if you apply some of the good advise above, like warping and different fonts between letters.

Attachment 9805

---

Hmm, yep, I think you're right. And, of course, OCR's are only getting better, so even if it's sufficient now, it won't be for long...

Anyway, thanks so much for running that through, Mario. Cheers!

Quote:

---

Originally Posted by Mario F.

Yeah, I was thinking about this too.

For illustration purposes, I ran the second picture in lab mode, extracted the lightness channel and applied a levels adjustment to it (reducing darkness and increasing the midtones). The image below was the result. Fairly easy to OCR. Now, my approach may not even be the easiest. But I think that by using lab mode, I guarantee I can tackle for any kind of noise in the image. As far as coding is concerned, I suppose this could be fairly easily achieved by someone bent on defeating this captcha.

---

That image can be even further cleaned up by a dilation/erosion:

Attachment 9806

Random noise can usually be removed. In this case I think it could be almost completely removed by calculating the variance of the chrominance. The noise is randomly distributed in chroma space, the signal is not.

Quote:

---

Originally Posted by brewbuck

That image can be even further cleaned up by a dilation/erosion:

Attachment 9806

Random noise can usually be removed. In this case I think it could be almost completely removed by calculating the variance of the chrominance. The noise is randomly distributed in chroma space, the signal is not.

---

Wow, yeah, I'm definitely going to need a more sophisticated algorithm! :)

Thanks again for all of the input, guys.

#1 is pretty hard on my Old eyes

#2 is good