WO2010028517A1

WO2010028517A1 - System and method for generating/ identifying cipher code via artificial neural network

Info

Publication number: WO2010028517A1
Application number: PCT/CN2008/001915
Authority: WO
Inventors: 陈淮琰; 王小春
Original assignee: 无敌科技(西安)有限公司
Priority date: 2008-09-09
Filing date: 2008-11-24
Publication date: 2010-03-18
Also published as: CN101350155A

Abstract

The system and method for generating/ identifying cipher code via artificial neural network dealing with the handwriting inputted from user are provided. The system involves a receiver module (110) for receiving a cipher code handwriting, an analyzing module (120) for analyzing the cipher code handwriting and outputting the cipher vector according to the result of the analyzing, a storage module (130) for storing an interference vector, and an artificial neural network (140) self-learning according to the interference vector and the cipher vector to generate the cipher weight value. The receiver module (110) is connected to the artificial neural network (140) via the analyzing module (120), and the storage module (130) is connected to the artificial neural network (140).

Description

Method and system for generating/verifying passwords by neural network

The present invention relates to a system and method for generating/verifying passwords, and more particularly to a system and method for generating/verifying passwords by processing a user-written handwriting through a neural network.

Background technique

Protecting one's own property is the nature of mankind. In order to prevent others from stealing private assets, from the password dial on the general safe to the computer's power-on password, e-mail login password, bank transaction transaction password, etc., password protection technology has been widely used in contemporary daily life.

Different password protection technologies have different advantages and disadvantages. In terms of character passwords which are widely used in various aspects at present, they are mostly composed of letters, numbers and other inputtable symbols, and have the characteristics of simple implementation, low cost and convenient use. . However, the user must remember the character as a password, which is a burden for people with poor memory. Moreover, since the password space of the character password is limited in complexity, the crack rate is quite high.

There is another password protection technology, which is built-in password on the IC card/magnetic card, so that the user can directly use the IC card/magnetic card to open the password-protected object without having to memorize the password. Although the IC card/magnetic card is very convenient to use, its disadvantage is that it is easy to lose and requires additional cost to make a card.

In addition, in order to improve the identification of personal identities, fingerprint technology has been developed in recent years. This technology has been gradually popularized in high-priced electronic products such as notebook computers and personal digital assistants (PDAs). . Compared with character passwords, fingerprint passwords are more complicated in space, and because fingerprints are unique characteristics of individuals, the reliability of fingerprint passwords is quite high. However, in order to use a fingerprint as a password, a special fingerprint input device is required, which significantly increases the cost and cannot be widely used in the parity product.

Another personally identifiable password protection technique is face recognition/retina recognition. This technology requires dedicated scanning equipment that is costly and there is currently no mature technology available to the general public.

Summary of the invention

The invention solves the problem that the password protection technology in the background art cannot be effectively and in a low cost manner. The technical problem of improving the crack rate of the password, and providing a method and system for generating/verifying the password through the neural network, and improving the reliability of the password under the premise of low cost.

The technical solution of the present invention is: The present invention provides a method for generating a password through a neural network, which is special in that: the method comprises the following steps: 1) receiving a password handwriting;

2) Analyze the password handwriting, and obtain the password vector according to the password handwriting;

Usually, the handwriting is a set of points that are represented by coordinates. There are several ways to treat this sample point as a set of vectors. Here are two examples:

2.1) According to the coordinates, the grid is divided into 100X100, and the sampling points in each grid are filtered into straight lines to obtain the slope of the line, and the slopes in all the grids are recorded as a set of vectors.

2.2) Divide all the sample points into segments, such as 50 segments, according to the number of sample points, and then filter each sample point into a straight line to obtain the slope of the line. Record the slope of all segments as a set of vectors;

3) providing an interference vector, inputting the cipher vector and the interference vector to the neural network;

4) The neural network is self-learning based on the cipher vector and the interference vector to generate a cryptographic weight value. The above neural network includes an input layer, an output layer, and a hidden layer, wherein the input layer has input neurons; the output layer has output neurons; the hidden layer has hidden neurons, and the input layer accesses the hidden layer through input neurons. The neurons are hidden, and the hidden layer accesses the output neurons of the output layer by hiding the neurons.

The specific steps of step 4 above are as follows:

4.1) the input neuron receives a cipher vector and an interference vector;

4.2) Calculating the cipher vector and the interference vector according to the hidden layer, inputting the cryptographic vector into the neural network that has been trained using the interference vector, and then outputting the operation result by each output neuron;

4.3) separately calculating an error value between the operation result of the output of each output neuron and the learning target value corresponding to the neural network;

The learning target value is specified in advance, for example, the interference vector specifies the output as a value between 000000 and 111110, and the cipher vector specifies the output as 1 1 1 1 1 1 . Feedback training-like neural networks are performed using these specified learning target values. 4.4) judging whether the error value is greater than the critical value, when the error value is not greater than the critical value, storing the operation result of each output neuron as a comparison value, and proceeding to the step; when the error value is greater than the critical value, proceeding to step 4.5);

The threshold can be specified by the user. When the critical value is extremely high, the handwriting height is required to be consistent to confirm. This value can usually be obtained empirically based on the specific implementation.

4.5) The neural network adjusts the weight values of the output layer and the hidden layer according to the error value (;

4.6) Obtaining the weight value in the neural network is the password weight value;

The weight value is a parameter set that determines the network function after the neural network training.

The method also includes the step of initializing a weight value of the neural network.

The present invention also provides a system for generating a password through a neural network, which is special in that: the system includes a receiving module for receiving a password handwriting, an analysis for analyzing the password handwriting, and outputting a password vector according to the analysis result. The module, the storage module for storing the interference vector, and the neural network based on the interference vector and the cryptographic vector to generate the cryptographic weight value; the receiving module accesses the neural network through the analysis module, and the storage module accesses the neural network.

The system also includes an initialization module for initializing the weight value of the neural network, the initialization module: an access type neural network.

The system also includes a readout module for reading an interference vector in the storage module, the storage module accessing the neural network via the readout module.

The present invention also provides a method for verifying a password through a neural network, which is special in that: the method comprises the following steps:

1) receiving the handwriting to be inspected;

2) Analyze the handwriting to be inspected and obtain the vector to be tested;

The method of obtaining the vector to be tested here is the same as the method of obtaining the cipher vector.

3) inputting the vector to be tested into the neural network, and the output neuron of the neural network outputs the value to be tested according to the to-be-detected vector;

4) The values to be checked are compared with the control values (the comparison values refer to the values specified by the cipher vectors in 4.3 of the foregoing steps), and the passwords are verified based on the comparison results.

The specific steps of step 4) above are as follows: 4.1) determining whether the maximum value in the value to be checked is less than the lower matching limit;

4.2) If the maximum value to be detected is less than the lower matching limit, proceed to step 4.5); if the maximum value to be detected is not less than the lower matching limit, proceed to step 4.3);

4.3) The neural network then determines whether the output neuron outputting the maximum value to be detected is the same as the output neuron outputting the maximum control value; if the two are the same neuron, proceed to step 4.4) if the two are different nerves Yuan, then proceed to step 4.5);

4.4) The password verification is successful;

4.5) Password verification failed.

The present invention also provides a system for verifying a password through a neural network, which is special in that: the system includes a receiving module for receiving a handwriting to be inspected, configured to analyze the handwriting to be inspected, and outputting a to be inspected according to the analysis result. An analysis module of the vector and a neural network for comparing the value to be checked with the comparison value; the receiving module accesses the neural network through the analysis module.

In summary, the present invention generates/verifies a password composed of handwriting through a neural network, and 'because the handwriting password is a two-dimensional directional trajectory, the space complexity is quite high and it is difficult to be cracked by the machine. In addition, the content of the handwriting password is memorized and recognized through the neural network. Even the person who knows the neural network's network data cannot reverse or crack the content of the handwriting password from the data of the neural network. The present invention has high security in comparison with the conventional technique of using characters as passwords.

In addition, the electronic product of the present invention only needs to have a general handwriting input device (such as a mouse, a tablet, etc.) for the user to input handwritten handwriting, and no additional device is required, so compared with the conventional password protection technology, The invention can generate a password with high security at a low cost.

In addition, for the user, the handwriting is only required to input the handwriting into the system of the present invention, and the password generation and verification can be performed. It is known to memorize the character password or carry the IC card/magnetic card to perform the password verification. In contrast, the present invention is more convenient to use.

DRAWINGS

1 is a schematic diagram of a specific embodiment of a system for generating a password by a neural network according to the present invention; FIG. 2 is a schematic diagram of a specific embodiment of a method for generating a password by a neural network according to the present invention; FIG. 3 is a schematic diagram of a neural network of the present invention. Schematic;

4 is a flow chart of a neural network-like self-learning of the present invention; FIG. 5 is a schematic diagram of a specific embodiment of a system for verifying a password by a neural network according to the present invention; FIG. 6 is a schematic diagram of a specific embodiment of a method for verifying a password by a neural network according to the present invention. detailed description

Referring to Figures 1 and 2, the system 100 for generating a password via a neural network of the present invention includes a receiving module 110, an analyzing module 120, a storage module 130, and a neural network 140. After the user inputs the password handwriting to be set as a password to the system 100 through a handwriting input device (not shown, such as a mouse, a tablet, etc.), the receiving module 110 is responsible for receiving the password handwriting (step 210), and the password handwriting is performed. Transfer to the analysis module 120. Next, the cryptographic handwriting is analyzed by the analysis module 120 (step 220), and the analysis module 120 of the embodiment includes a sampling unit 122 for sampling the cryptographic handwriting to obtain a cryptographic vector.

The storage module 130 is configured to store at least one interference vector. In the present embodiment, the interference vector is obtained, for example, by analysis of the interference handwriting previously input to the system 100 via the analysis module 120. The cipher vector and the interference vector are respectively composed of a plurality of vector data, for example. After the memory module 130 provides at least one interference vector (step 230), the cryptographic vector and the interference vector are input to the neural network 140 (step 240). In this embodiment, the system 100 further includes a readout module 150 for reading the interference vector in the storage module 130, and outputting the read interference vector together with the password vector output by the analysis module 120 to the class. Neural network 140.

It should be noted that the system 100 of this embodiment may further include an initialization module 160 for initializing the weight values in the neural network based on the input of the interference vector and the cryptographic vector to the neural network 140 (step 245). For example, initialization module 160, for example, initializes the weight value of neural network 140 to a random value between 0 and 1 in step 245.

Referring to FIG. 3, the neural network 140 includes an input layer 142, an output layer 144, and at least one hidden layer 146, wherein the input layer 142 has a plurality of input neurons 142a; the output layer has a plurality of output neurons 144a; and the hidden layer 146 has Multiple hidden neurons 146a. It is worth mentioning that the number of layers of the hidden layer 146 may vary depending on the complexity requirements for the password handwriting. For example, if a simple handwriting is to be used as the password, the number of layers of the hidden layer 146 can be reduced; if more complex handwriting is to be used as the password to encrypt the important data, the number of layers of the hidden layer 146 can be increased to provide more Complex recognition capabilities. Those skilled in the art can determine the number of layers of the hidden layer 146 according to their needs. The present invention is not This limits the number of layers of the hidden layer 146.

After receiving the cipher vector CV and the interference vector DV, the neural network 140 performs self-learning according to the cipher vector CV and the interference vector DV to generate a plurality of cryptographic weight values (step 250), and the neural network 140 remembers the passwords. The weight value (step 260) is used to facilitate subsequent verification of the password.

Referring to Fig. 4, after the cipher vector CV and the interference vector DV are input to the input layer 142 of the neural network 140, the cryptographic vector CV and the interference vector DV are received by the input neuron 142a (step 251). In the present embodiment, the number of input neurons 142a is, for example, equal to the sum of the amounts of vector data of the cipher vector CV and the interference vector DV. That is, each input neuron 142a receives a vector of data.

Each output neuron 144a outputs an operation result according to the operation of the cryptogram vector CV and the interference vector DV received by the input neuron 142a by the hidden layer 146 (step 252), and then respectively calculates the operation result 0 of each output neuron corresponding thereto. The error value between the learning target values (step 253). For example, the error value 5 between the kth output neuron and its corresponding learning target value is E _k = (T _k - O _k ) x O _k (l - O _k ) _} where _Tk is the first output The learning target value corresponding to the neuron, O _k is the operation result of the actual output of the kth output neuron.

Thereafter, it is judged whether or not the error value is greater than a critical value (step 254). When the error value between the operation result of each output neuron 144a and the corresponding learning target value is not greater than the threshold value, the operation result of each output neuron is stored as a comparison value (step 255), so that the password is subsequently performed. When verifying, these comparison values are used as a comparison benchmark. Moreover, the weight value in the neural network 140 at this time is the cryptographic weight value.

On the contrary, when the error value between the operation result of the output neuron 144a and the corresponding learning target value is greater than the critical value, the neural network 140 adjusts the weight values of the output layer 144 and the hidden layer 146 by itself. For example, when the error value E _k between the operation result O _{k of} the kth output neuron and the learning target value T _k is greater than the critical value, the kth output neuron is connected according to the error value E _k The weight value between the jth hidden neurons of the mth hidden layer (step 256).

In this embodiment, step 256 is, for example, a public To adjust the weight value between the kth output neuron and the jth hidden neuron of the mth hidden layer to which it is connected. among them, ^W ' ^k is the weight value after adjusting n+1 times between the kth output neuron and the jth hidden neuron of the mth hidden layer to which it is connected, which is the kth output neuron and the mth connected thereto The weight value after nth adjustment between the jth hidden neurons of the layer hidden layer. Where n is a positive integer; Oj is the result of the operation output by the jth output neuron; L is called the learning rate and is a constant greater than zero, which is usually between 0 and 1. In the self-learning process of the neural network, if the learning rate is large, the learning speed is faster; if the learning rate is smaller, the learning result is more accurate.

Next, an output error value Ej of the j-th hidden neuron of the m-th hidden layer is calculated according to the adjusted weight value in step 256 (step 257). In the present embodiment, the error value is calculated Ε, - Ο ^ - Ο χ ^ Ε ^ where ^ _(to the right after the step 256 by adjusting the weight values.

Then, according to the output error value Ej calculated in step ft257, the weight value between the jth hidden neuron and the i-th hidden neuron of the m-1th hidden layer to which it is connected is adjusted (step 258).

In this embodiment, step 258 adjusts the i-th hidden neuron of the hidden layer of the m-1th hidden layer and the m-1 layer hidden by the hidden neuron, for example, by the formula ⁼ ^'^><><''^>< The weight value between . ^ ' ⁺ ' . is the adjustment between the jth hidden neuron and the i-th hidden neuron of the m-1 hidden layer to which it is connected.: Weight value after n+1 times , the weight value adjusted n times after the jth hidden neuron and the i-th hidden neuron of the m-1th hidden layer to which it is connected, where m and n are positive integers, and !!^ .

Thus, step 256 to step 258 are continuously repeated until the error value between the operation result of each output neuron 144a and the corresponding learning target value is not greater than the critical value, and the neural network 140 completes its self-learning. The weight value memorized by the network 140 is the aforementioned password weight value.

In order to make the present invention more familiar with the present invention, the system and method for verifying the password generated by the above method and system will be described in detail below with reference to the drawings.

Referring to Figures 5 and 6, system 500 includes a receiving module 510, an analysis module 520, and a neural network 530. The receiving module 510 and the analyzing module 520 are similar to the receiving module 110 and the analyzing module 120 of the foregoing embodiment, respectively. In particular, the neural network 530 has self-learned according to the flow of steps shown in "Fig. 4", thus memorizing the cryptographic weight values described above.

After the user inputs the to-be-detected handwriting to the password verification system 500 by using a handwriting input device (not shown), the receiving module 510 is responsible for receiving the to-be-tested handwriting (step 610), and transmitting the to-be-detected handwriting to the analysis module 520. . Next, the handwriting to be detected by the analysis module 520 is analyzed (step 620). Its The analysis module 520 includes a sampling unit 522 for sampling the to-be-tested handwriting to obtain a to-be-checked vector. Like the cipher vector and the interference vector of the foregoing embodiment, the vector to be detected may also be composed of a plurality of vector data.

Thereafter, the test vector is input to the neural network 530 such that each output neuron of the neural network 530 outputs a test value according to the vector to be tested (step 630). The neural network 530 will automatically compare these values to the control values stored therein. In this embodiment, the step of comparing the value to be checked with the comparison value includes determining whether the maximum value of the values to be detected is less than a lower matching limit (step 642), wherein the lower limit of the matching is a constant, which is familiar to the skilled person. The matching accuracy determines the actual value of the matching lower limit.

If the result of the determination in step 642 is that the maximum to-be-checked value is less than the matching lower limit, it indicates that the password verification fails (step 648); if the result of the determination in step 642 is that the maximum to-be-checked value is not less than the matching lower limit, then The neural network 530 then determines whether the output neuron outputting the maximum test value is the same as the output neuron outputting the maximum control value (step 644). If the two are the same neuron, it means that the handwriting to be detected by the user matches the originally set password handwriting, that is, the password verification is successful (step 646). On the other hand, if the two are different neurons, it means that the handwriting to be detected by the user does not match the originally set password handwriting, and the password verification fails (step 648).

Claims

Claim

A method for generating a password by a neural network, characterized in that: the method comprises the following steps:

1) Receive password handwriting;

4) The neural network is self-learning based on the cipher vector and the interference vector to generate a cryptographic weight value.

2. The method for generating a password by a neural network according to claim 1, wherein: said neural network comprises an input layer, an output layer and a hidden layer, wherein the input layer has input neurons; and the output layer has output nerves. The hidden layer has hidden neurons that access the hidden neurons of the hidden layer through the input neurons, and the hidden layer accesses the output neurons of the output layer by hiding the neurons.

3. The method for generating a password by a neural network according to claim 2, wherein: the specific steps of the step 4) are as follows:

4.1) the input neuron receives a cipher vector and an interference vector;

4.4) judging whether the error value is greater than the critical value, when the error value is not greater than the critical value, storing the operation result of each output neuron as a comparison value, and proceeding to the step; when the error value is greater than the critical value, proceeding to step 4.5);

4.5) The neural network adjusts the weight values of the output layer and the hidden layer according to the error value;

4.6) The weight value in the class-like neural network is the cryptographic weight value.

4. A method of generating a password by a neural network according to claim 1 or 2 or 3, wherein: the method further comprises the step of initializing a weight value of the neural network.

5. A system for generating a password through a neural network, the system comprising: a receiving module for receiving a password handwriting, for analyzing the password handwriting, and outputting a password according to the analysis result. An analysis module of the vector, a storage module for storing the interference vector, and a neural network based on the interference vector and the cryptographic vector to generate a cryptographic weight value; the receiving module accesses the neural network by analyzing the modulo block The storage module is connected to a neural network.

A system for generating a password by a neural network according to claim 5, wherein: the system further comprises an initialization module for initializing a weight value of the neural network, the initialization module accessing the neural network.

7. The system for generating a password by a neural network according to claim 5 or 6, wherein: the system further comprises a readout module for reading an interference vector in the storage module, the storage module being read by The module accesses a neural network.

8. A method for verifying a password by a neural network, the method comprising: the method comprising the steps of:

1) receiving the handwriting to be inspected;

2) Analyze the handwriting to be inspected and obtain the vector to be tested;

3) inputting the vector to be tested into the neural network, and outputting the neuron of the neural network to output the value to be tested according to the vector to be detected;

4) Compare these values to the control values and verify the password based on the comparison results.

9. The method for verifying a password by a neural network according to claim 8, wherein: the specific steps of the step 4) are as follows:

4.1) Determine whether the maximum value in the value to be checked is less than the lower matching limit;

4.4) The password verification is successful;

4.5) Password verification failed.

10. A system for verifying a password by a neural network, wherein: the system includes a receiving module for receiving a handwriting to be inspected, configured to analyze the handwriting to be inspected, and outputting a pending test according to the analysis result. An analysis module of the vector and a neural network for comparing the value to be checked with the comparison value; the receiving module accesses the neural network through the analysis module.