12306 Graphical Verification Code Analysis

Verification code is a very interesting question, its purpose is to distinguish the input is a person or machine, the problem is essentially a Turing test (recommended film "Imitation Game"), the verification code is a simple and efficient verification method, by the professor of the CMU in 2000 created around, Later this cattle people will be scattered verification code collected, converted into a huge productivity, the success of the articles paper-based article digitization, Google also used its identification number, road signs and so on (a man of God created a verification code, Also made a significant contribution to the verification code). 12306 yesterday changed to a graphics verification code, and in fact, the graphics verification code is not new, as early as a few liken,Google replaced by a graphical verification (Google to make verification code more simple), the picture is as follows:

Take a look at our 12306 Image Verification code:

Note: If you choose the wrong, 2s, you can try again, and the picture has been changed, but you can try N times.

Several features of 12306 graphics verification Code

1. Double verification of text + image

Graphics this is a lot of features, after the elaborate. Text This relative to the simple text verification code, the difficulty coefficient decreased a lot, basically only font, bold, size, position changes. Generally speaking, for this kind of non-distortion (easy to identify), overlapping small (easy to split), the background of a single (easy to binary), the Chinese printing (easy to identify), the recognition algorithm is very mature, this, I believe we have seen the OCR software powerful recognition function. In addition, the text here is not isolated, it must be a noun, which can further improve the recognition rate. Therefore, it is relatively easy to crack this off.

2. Pictures are familiar

such as greeting cards, statues, shells, glass bottles, rolling pins, hot air balloons, and so on, most of the daily life can see. This determines that there are not too many categories. At present, the total number of ImageNet (the largest database of image recognition currently) is 21841 categories, taking into account the common , 12306 of the category should not be higher than this number (I think it should be far below this number), the possibility of increasing the number of categories in the future is not very , which makes the classification of the problem not particularly large , but the expansion of the number of each category is inevitable .

3. The database is taken from the Internet

12306 of the data tagged from the Internet, that is, for "basketball", 12306 may be directly inside Google/Baidu search "basketball", and then add the resulting image to the database. The reason for this is two points. First of all, 12306 requires a huge amount of tag data, if the database is too small, then the cracked software can collect this data, to compare. This is like we got the teacher's question bank, no matter how the teacher, I can calmly deal with. Second, 12306 of the cost of self-tagging is too large, considering that the database needs to be constantly updated, and future maintenance costs are too high. At present, for the research of the field of image detection, there are a lot of database markers directly taken from the Internet. The use of Internet pictures and marks is convenient, however, this will also bring convenience to crack, so-called "analyticals Xiao", because you can directly use the Internet data to help crack . Here are two chestnuts:

Google Image Search

For example, the ticket software can search for images in Google, then get the title of the link, and then match the title with the text to be found, such as the "Spring Festival couplets" in the example above, which becomes a small problem of text search , do not need to identify. Of course, this practice of network traffic and time consumption is too large (close to 1s), this ticket-snatching software is built, it is difficult to apply. Therefore, the ticket-grabbing software may be more inclined to localize the library .

Baidu Search "Handcuffs"

Since the categories are not many and the categories are fixed, the hack software can also use tags to search for images on the internet and then create its own image Library . This idea for 12306来 said, personally think the threat is very big! , the image is exactly matched. Even if the library to rob the software is not 12306 comprehensive, and even with 12306 contains the image inconsistent, there is no big relationship, not big, cracked software for each type of training a classifier, such as specialized training a handcuff classifier, and then to the 8 picture classification is good, generally speaking, two classification The classification effect is very accurate. Of course, you can also directly use a multi-classifier , such as CNN/DNN, now deep learning of the classification results can also, here to everyone a link. However, the exact classification of the idea to do, the accuracy will certainly not reach, because it requires the computer has the same visual and conceptual understanding of the human brain.

4. Generally, only choose two

Supposedly, in order to increase the uncertainty as much as possible, the same group of the number of the same class can take 1,2,3,....,8, but in fact, considering the user feel, the number should not exceed 3, otherwise users will dozens ..., even if 3, many people will not be happy, and a too little, guess 8 times, 2 There are 56 kinds of situations, or more, so I guess most of the cases are only 2 belong to the same class, and very few cases there are only 1 or 3, I believe in the future how to change, this setting should be unchanged. This has brought a lot of convenience to the crack, because most of the time, when the crack does not need to know what category to look for, do not need to recognize the text, just need to find the two pictures of the most similar picture can be, so this translates into a similarity matching problem ( Many methods, such as the simplest gray histogram matching, gray enough to add gradient direction ah, consider the rotation of the Brahma direction ah ....).

5. The creation of each group of images may use similarity

For example, the thermos is similar to a sedan chair, and the cell phone shell and the group have a lot of rectangles, which, obviously, can cause a little disturbance to the similarity and recognition.

6. There may be multiple classes in each group with more than one picture

For example, the first picture, in addition to the charger, there are 2 suits, the second picture in addition to the screwdriver, there are 3 of mooncakes. Obviously, this is to deal with the 4th limitation of the interference, it is obvious that such interference does not bring much difficulty , for example, the first one has 2 categories have 2, then each guessed the probability of 50%, the big deal has 4 classes, each class 2, the probability is also up to 25%. Although the use of 4 categories of uncertainty is greater, but this special set of conditions, but also to facilitate the identification, so there is generally only 1-2 classes (this part is not rigorous).

Today, 360 said they have successfully cracked, and the success rate of the ticket has increased by 200% ..., the speed of the fast, raspberry, hey, however persuasive, outsmart Ah! However, 12306 is just the use of graphics verification code, this method there are many can be adapted to the place ( such as the appropriate rotation of the picture), in short, 12306 there are many tricks can be used, rob the ticket software can catch the first move, does not mean that the follow-up tricks can easily catch, and, I guess rob the software this time to take some opportunistic , really to solve this problem, it is necessary to use machine learning (Baidu, Google is essentially machine learning), this is a research and need long-term research problems, otherwise LSVRC (Large Scale Verification Code recognition competition) Such challenges can be closed.

Branching out a sentence, anyway, 12306 to promote Chinese characters and common sense made a great contribution, for example, I did not know the arch long what kind of, which reminds me of my 5-year-old into the preschool scene, the teacher thought I was too young to test my IQ, So take out a book, I want to point out which is the tiger, which is the elephant ..., now think this is where the IQ ah, clearly see me is not a love to watch cartoon children!

12306 Graphical Verification Code Analysis