A privacy-preserving approach to social media fake news detection

Abstract

The present invention discloses a privacy-preserving social media fake news detection method, which uses text fingerprints and locally sensitive hashing technology to convert fake message text into a keyword vector that can be quickly compared for similarity, and then uses searchable encryption technology to perform keyword search on a ciphertext data set, ultimately achieving comparison and determination of ciphertext fake messages, and the detection speed is fast and the robustness is strong. In addition, the present invention uses an inadvertent pseudo-random function to ensure that the user's query information will not be leaked to a third-party fake news identification agency and a social media fake news detector while determining whether the message is a fake message. The present invention can also control the number of services through encryption authorized by a third-party fake news identification agency, further ensuring that the identified fake message list released by the fake news identification agency will not be maliciously tracked or tampered with by a social media platform.

Description

Technical Field

[0001] The present invention belongs to the technical field of network security, and in particular relates to a method for detecting fake news on social media. Background Art

[0002] Social media mainly includes social networking sites, Weibo, WeChat, forums, etc. After more than 10 years of development, China has become the world's largest mobile social media market, with more than 1 billion social media users. With the rapid expansion of the market size, the social interaction and information dissemination functions carried by mobile social media are also constantly innovating, affecting and changing all aspects of interpersonal interaction, working methods, and transaction services. We can now easily obtain information and request services through mobile social software. However, while enjoying the convenience brought by social media, social media platforms have also become the main way to spread false information. Under the temptation of huge economic benefits, there is even a phenomenon of maliciously creating false information and slandering others. In addition, the harm caused by false information cannot be ignored. At the individual level, false information not only misleads people's judgment and disrupts people's thinking, but also makes it difficult for people to distinguish between right and wrong and true and false. Under the wrong guidance of false information, some people are likely to believe in false information, affecting their personal value judgments and choices. At the national level, the bad public opinion orientation brought about by false information will affect the formulation of national policies, have a negative impact on public interests, and even cause social order unrest.

[0003] In response to this type of problem, fake news detection technology has emerged, which aims to help users better identify the authenticity of news when browsing it. Currently, this field has been widely researched and applied in both industry and academia. The mainstream fake news detection in the industry mainly uses manual or algorithmic detection methods. For example, YouTube has cooperated with medical experts worldwide to reduce low-quality video news about COVID19 on the platform. WeChat has developed a small program called "WeChat Rumor Refutation Assistant", which obtains the user's reading history and compares it with the collected fake news dataset to detect fake news. The academic community has also made public many datasets on fake news on social media, and proposed many machine learning methods to detect and locate fake news by analyzing the grammar and semantic characteristics of text messages, or user dynamics (likes, follows, and reposts).

[0004] Although the current fake news detection mechanism on social media can suppress the spread of fake news to a certain extent, the mainstream detection methods are mostly based on text content information. Therefore, when users use this service, they often worry that social media will collect and abuse their browsing history privacy information, or even conduct background checks or behavior monitoring on specific users. Therefore, it is necessary to design a new method through cryptographic methods that can effectively protect user privacy and effectively block the spread of fake news. At the same time, it can also encourage more fake news identification agencies to provide more fake news data sets in different scenarios. Summary of the invention

[0005] In order to overcome the deficiencies of the prior art, the present invention provides a privacy-preserving social media fake news detection method, which uses text fingerprints and locally sensitive hashing technology to convert fake message text into a keyword vector that can be quickly compared for similarity, and then uses searchable encryption technology to perform keyword searches on ciphertext data sets, ultimately achieving comparison and determination of ciphertext fake messages, and with fast detection speed and strong robustness. In addition, the present invention uses an inadvertent pseudo-random function to ensure that the user's query information will not be leaked to a third-party fake news identification agency and a social media fake news detector while determining whether the message is a fake message. The present invention can also control the number of services through encryption authorized by a third-party fake news identification agency, further ensuring that the list of identified fake messages released by the fake news identification agency will not be maliciously tracked or tampered with by the social media platform.

[0006] The technical solution adopted by the present invention to solve the technical problem includes the following steps:

[0007] Step 1: The third-party fake news identification agency initializes system parameters;

[0008] First, we construct an oblivious pseudo-random function EC-OPRF based on the elliptic curve discrete logarithm problem ECDLP: H(k,m) = {0,1} l ×{0,1} * →{0,1} l , and generate a random key pair (k 1 ,k 2 ), where the key k 1 ,k 2 are all numbers of length l, that is, {0,1} l , the message m is a string of indefinite length, that is, {0,1} * , and initialize the parameters of Fingerprint and locality sensitive hashing LSH functions;

[0009] Step 2: A third-party fake news identification agency releases the latest fake news collection B;

[0010] For each identified fake message mi ∈B, first normalize the fake message text, then calculate the fake message text fingerprint and LSH keyword v i And add it to the list V, that is, V = {v 1 ,…,v n};

[0011] Step 3: A third-party fake news identification agency generates an encrypted fake news dataset D;

[0012] Traverse the list V obtained in step 2, for each component v in v i , using k 1 Encrypt v, k 2 Encrypted fake message i The encrypted result is added as a key-value pair into the SSE framework, and finally the encrypted dataset D is obtained. The dataset D is then synchronized to the social media fake news detection provider.

[0013] Step 4: The user-side program normalizes the user query plaintext information m q ;

[0014] The user installs and uses the service program provided by the fake news identification agency. When the user has doubts about the authenticity of the message and needs to verify it, the user service program reads the plain text information of the message and normalizes it to obtain m q ;

[0015] Step 5: The user service program generates a blinded query request x and sends it to a third-party fake news identification agency to request authorization;

[0016] The user service program calculates m using the same algorithm as in step 2 q The text fingerprint and LSH keyword v q ; Then the user service program generates a random number r as the blinding factor and caches it locally. q Each component v q [i] Calculate the blind request x[i] = mask(r,v q [i]), mask(.) indicates blinding, and the blinded query request x is sent to a third-party fake news identification agency;

[0017] Step 6: The third-party fake news identification agency generates an authorized blind query request and returns it to the user service program;

[0018] For the blinded query request x obtained in step 5, use the key pair (k 1 ,k 2 ) to perform encryption authorization; for each component of x, calculate the encryption authorization y 1 [i] = k1 x[i],y 2 [i] = k 2 x[i], and the encrypted authorized result (y 1 ,y 2 ) Return to the user service program;

[0019] Step 7: The user service program removes the blinding factor to obtain the authorized query information (t 1 ,t 2 );

[0020] For the encrypted authorization result (y 1 ,y 2 ), use the blinding factor r cached in step 5 to perform a deblinding operation to obtain the authorized query information (t 1 ,t 2 ), where t 1 [i] = unmask(r -1 ·y 1 [i]), t 2 [i] = unmask(r -1 ·y 2 [i]), unmask(.) means unblinding, r -1 It is obtained by solving the modular inverse operation of the blinding factor r with respect to the order of the elliptic curve;

[0021] Step 8: The user service program sends the authorized query information to the social media fake news detection provider for encrypted search, and counts the number of matching LSH keywords in the encrypted data set to derive the probability that the message is fake, and finally returns the query result to the user.

[0022] Preferably, the EC-OPRF function H(k,m)={0,1} l ×{0,1} * →{0,1} l There are three hash functions H 1 :{0,1} * →{0,1} l , H 2 :{0,1} * →E,H 3 :E→{0,1} l Realization, that is, H(k,m)=H 3 (k·H 2 (H 1 (m))), where E is a point on the elliptic curve, H 1 is the cryptographic hash function SHA256, the hash function H 2 and H 3 Implemented through the Elliptic Curve Cryptography library.

[0023] Preferably, the elliptic curve is a NIST P-192 elliptic curve.

[0024] Preferably, the text fingerprint algorithm is implemented using the MinHash algorithm.

[0025] Preferably, the locality sensitive hash (LSH) function is implemented using a Hamming-LSH algorithm.

[0026] Preferably, the blinding mask (.) uses elliptic curve multiplication operation.

[0027] The beneficial effects of the present invention are as follows:

[0028] The present invention can realize the comparison and judgment of encrypted false news, and has fast detection speed and strong robustness; the present invention can also control the service times by means of encryption authorized by a third-party false news identification agency, and further ensure that the identified false news list released by the false news identification agency will not be maliciously tracked or tampered with by the social media platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A scenario graph for deploying the social media fake news detection service of the present invention.

[0030] Figure 2 The detailed process of false information detection of the present invention is shown. DETAILED DESCRIPTION

[0031] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0032] In order to overcome the shortcomings of the prior art, the present invention proposes a privacy-preserving social media fake news detection method. By using text fingerprints and locally sensitive hashing technology, the fake message text is converted into a keyword vector that can be quickly compared for similarity, and then the searchable encryption technology is used to perform keyword searches on the ciphertext data set, and finally the comparison and judgment of the ciphertext fake message is achieved, and the detection speed is fast and the robustness is strong. In addition, the present invention uses an inadvertent pseudo-random function and a searchable symmetric encryption framework to determine whether the message is a fake message while ensuring that the user's query information will not be leaked to a third-party fake message identification agency and a social media fake message detector. The present invention can also control the number of services through encryption authorized by a third-party fake message identification agency, further ensuring that the list of identified fake messages released by the fake message identification agency will not be maliciously tracked or tampered with by the social media platform.

[0033] A privacy-preserving social media fake news detection method comprises the following steps:

[0034] Step 1: The third-party fake news identification agency initializes system parameters;

[0035] First, we construct an oblivious pseudo-random function EC-OPRF based on the elliptic curve discrete logarithm problem ECDLP: H(k,m) = {0,1} l ×{0,1} * →{0,1} l , and generate a random key pair (k 1 ,k 2 ), where the key k 1 ,k 2 are all numbers of length l, that is, {0,1} l , the message m is a string of indefinite length, that is, {0,1} * , and initialize the parameters of Fingerprint and locality sensitive hashing LSH functions;

[0036] Step 2: A third-party fake news identification agency releases the latest fake news collection B;

[0037] For each identified fake message m i ∈B, first normalize the fake message text, then calculate the fake message text fingerprint and LSH keyword v i And add it to the list V, that is, V = {v 1 ,…,v n};

[0038] Step 3: A third-party fake news identification agency generates an encrypted fake news dataset D;

[0039] Traverse the list V obtained in step 2, for each component v in v i , using k 1 Encrypt v, k 2 Encrypted fake message i The encrypted result is added as a key-value pair into the SSE framework, and finally the encrypted dataset D is obtained. The dataset D is then synchronized to the social media fake news detection provider.

[0040] Step 4: The user-side program normalizes the user query plaintext information m q ;

[0041] The user installs and uses the service program provided by the fake news identification agency. When the user has doubts about the authenticity of the message and needs to verify it, the user service program reads the plain text information of the message and normalizes it to obtain m q ;

[0042] Step 5: The user service program generates a blinded query request x and sends it to a third-party fake news identification agency to request authorization;

[0043] The user service program calculates m using the same algorithm as in step 2 q The text fingerprint and LSH keyword v q ; Then the user service program generates a random number r as the blinding factor and caches it locally. q Each component v q [i] Calculate the blind request x[i] = mask(r,v q [i]), mask(.) indicates blinding, and the blinded query request x is sent to a third-party fake news identification agency;

[0044] Step 6: The third-party fake news identification agency generates an authorized blind query request and returns it to the user service program;

[0045] For the blinded query request x obtained in step 5, use the key pair (k 1 ,k 2 ) to perform encryption authorization; for each component of x, calculate the encryption authorization y 1 [i] = k 1 x[i],y 2 [i] = k 2 x[i], and the encrypted authorized result (y 1 ,y 2 ) Return to the user service program;

[0046] Step 7: The user service program removes the blinding factor to obtain the authorized query information (t 1 ,t 2 );

[0047] For the encrypted authorization result (y 1 ,y 2 ), use the blinding factor r cached in step 5 to perform a deblinding operation to obtain the authorized query information (t 1 ,t 2 ), where t 1 [i] = unmask(r -1 ·y 1 [i]), t 2 [i] = unmask(r -1 ·y 2 [i]), unmask(.) means unblinding, r -1 It is obtained by solving the modular inverse operation of the blinding factor r with respect to the order of the elliptic curve;

[0048] Step 8: The user service program sends the authorized query information to the social media fake news detection provider for encrypted search, and counts the number of matching LSH keywords in the encrypted data set to derive the probability that the message is fake, and finally returns the query result to the user. Specific embodiment:

[0050] See also Figure 1 , Figure 2 As shown, the specific steps of the present invention are as follows:

[0051] 1. The third-party fake news identification agency initializes the system parameters. First, based on the Elliptic Curve Discrete Logarithm Problem (ECDLP), an oblivious pseudo-random function EC-OPRF function H(k,m) is constructed to generate a random key pair (k 1 ,k 2 ), and initialize the parameters of the Fingerprint and LSH functions, and finally initialize the three hash functions H 1 :{0,1} * →{0,1} l , H 2 :{0,1} * →E,H 3 :E→{0,1} l , where E is a point on the elliptic curve, and {0,1} * and {0,1} l They represent a string of indefinite length and a string of length l respectively.

[0052] Specifically, the hash function H 1 You can use the common cryptographic hash function SHA256, the hash function H 2 and H 3 You need to use the elliptic curve cryptography library to implement it. The elliptic curve uses the NIST P-192 elliptic curve.

[0053] 2. A third-party fake news identification agency identifies and publishes the latest fake news collection B. For each identified fake news m i ∈B, first normalize the text, then calculate the text fingerprint and LSH keyword v i , add them to the list V, that is, V = {v 1 ,…,v n}.

[0054] Specifically, in order to increase the accuracy of detection, we perform normalization processing on each fake message by removing common words and symbols, and then use the MinHash algorithm to hash the high-dimensional text data to obtain a text fingerprint. Then, we use the LSH hash function based on the Hamming distance to further convert the text fingerprint into an LSH keyword vector.

[0055] 3. A third-party fake news identification agency generates an encrypted fake news dataset D. Traverse the list V obtained in step 2, and for each component v in v i , using k 1 Encrypt v, k 2 Encrypted fake message i The encrypted result is added as a key-value pair into the SSE framework. Finally, the encrypted dataset D is obtained. The dataset D is synchronized to the social media fake news detection provider.

[0056] Specifically, the fake news identification agency maintains an auxiliary key-value data structure And traverse each v in step 2 i ∈V. For each component v in v i , using the key pair (k 1 ,k 2 ), encrypt the 1 =H 3 (k 1 ·H 2 (H 1 (v))),t 2 =H 3 (k 2 ·H 2 (H 1 (v))), where the symbol · represents the multiplication operation on the elliptic curve. For each calculated t 1 ,t 2 , if available in Find the key t 1 The corresponding value n, then let c = n + 1 and (t 1 ,c)Replace(t 1 ,n)Insert If in Key t not found 1 The corresponding value n, then let c = 1, and (t 1 ,1)Insert Then use the function F and t 1 Calculate the authentication code of message c, that is, a = F(t 1 ,c), in the present invention, function F is implemented by using HMAC-SHA1 signature authentication algorithm. The identifier i of the fake message is encrypted using the symmetric encryption function Enc, that is, b = Enc (t2 ,i), and finally insert (a,b) as a key-value pair into the encrypted data set D. In the present invention, the function Enc is implemented using the AES encryption algorithm. .

[0057] 4. User-side program normalization User query plain text information m q The user installs and uses the service program provided by the fake news identification agency. When the user has doubts about the authenticity of a message and needs to verify it, the user service program reads the plain text information of the message and normalizes it to obtain m q .

[0058] Specifically, step 4 uses the same normalization method as step 2.

[0059] 5. The user service program generates a blinded query request x and sends it to a third-party fake news identification agency for authorization. The user service program further calculates m using the algorithm in step 2 q The text fingerprint and LSH keyword v q . The user service program then generates a random number r as the blinding factor and caches it locally. q Each component v q [i] Calculate the blind request x[i] = mask(r,v q [i]) = r·H 2 (H 1 (v q [i])) and sends the blinded query request x to a third-party fake news identification agency.

[0060] Specifically, the blinding operation generally uses elliptic curve multiplication to make it difficult for others to calculate v without r. q [i].

[0061] 6. The third-party fake news identification agency generates an authorized blind query request and returns it to the front-end user service program. For the blind query request x obtained in step 5, use the key pair (k 1 ,k 2 ) is used for encryption authorization. For each component of x, the encryption authorization y is calculated separately. 1 [i] = k 1 ·x[i]y 2 [i] = k 2 x[i]. And the encrypted authorized result (y 1 ,y 2 )Return to the user service program.

[0062] 7. The user service program removes the blinding factor to obtain the authorized query information (t 1 ,t2 ). For the encrypted authorization result (y 1 ,y 2 ), use the blinding factor r cached in step 5 to perform a deblinding operation to obtain the authorized query information (t 1 ,t 2 ), t 1 [i] = unmask(r -1 ·y 1 [i])t 2 [i] = unmask(r -1 ·y 2 [i]).

[0063] Specifically, here r -1 It is obtained by performing a modular inverse operation on the blinding factor r in step 5 with respect to the order of the elliptic curve.

[0064] 8. The user service program sends the authorized query information to the social media fake news detection provider for encrypted search, and counts the number of matching LSH keywords in the encrypted data set to derive the probability that the message is false, and finally returns the query result to the user.

[0065] Specifically, for t 1 Each component t 1 [i] Social media fake news detectors all maintain a counter c, whose initial value is 0. 1 [i] First, use the function F(t 1 [i],c) Calculate the query item o. If o can be found in the dataset D, use t 2 [i] Decrypt to get the fake message identifier u id = Dec(t 2 [i], D.get(o)), and increment c by 1, recalculate the query item o, and re-query D until no such o can be found in D. Then i+1 is added, and the above process is repeated until the end. Finally, the number of matched LSH keywords is counted, the probability that the message is a false message is calculated, and the result is returned to the user service program, and a warning message is sent to the user.

Claims

1. A privacy-preserving social media fake news detection method, characterized in that: The steps include: Step 1: The third-party fake news identification agency initializes system parameters; First, we construct an oblivious pseudo-random function EC-OPRF based on the elliptic curve discrete logarithm problem ECDLP: H(k,m) = {0,1} l ×{0,1} * →{0,1} l , and generate a random key pair (k1, k2), where the keys k1 and k2 are both numbers of length l, i.e. {0, 1} l , the message m is a string of indefinite length, that is, {0,1} * , and initialize the parameters of Fingerprint and locality sensitive hashing LSH functions; Step 2: A third-party fake news identification agency releases the latest fake news collection B; For each identified fake message m i ∈B, first normalize the fake message text, then use the MinHash algorithm to calculate the text fingerprint, and use the LSH hash function based on Hamming distance to further convert the text fingerprint into the LSH keyword v i , and add them to the list V, that is, V = {v1,…,v n }; Step 3: A third-party fake news identification agency generates an encrypted fake news dataset D; Traverse the list V obtained in step 2, for each component v inv i , use k1 to encrypt v, k2 to encrypt fake message m i The encrypted result is added as a key-value pair into the SSE framework, and finally the encrypted dataset D is obtained. The dataset D is then synchronized to the social media fake news detection provider. Step 4: The user-side program normalizes the user query plaintext information m q ; The user installs and uses the service program provided by the fake news identification agency. When the user has doubts about the authenticity of the message and needs to verify it, the user service program reads the plain text information of the message and normalizes it to obtain m q ; Step 5: The user service program generates a blinded query request x and sends it to a third-party fake news identification agency to request authorization; The user service program calculates m using the same algorithm as in step 2 q The text fingerprint and LSH keyword v q ; Then the user service program generates a random number r as the blinding factor and caches it locally. q Each component v q [i] Calculate the blind request x[i] = mask(r,v q [i]), mask(.) indicates blinding, and the blinded query request x is sent to a third-party fake news identification agency; Step 6: The third-party fake news identification agency generates an authorized blind query request and returns it to the user service program; For the blinded query request x obtained in step 5, use the key pair (k1, y2) stored locally in step 1 to perform encryption authorization; for each component of x, calculate the encryption authorization y1[i] = k1·x[i], y2[i] = k2·x[i], and return the encrypted authorization result (y1, y2) to the user service program; Step 7: The user service program removes the blinding factor to obtain the authorized query information (t1, t2); For the encrypted authorization result (y1, y2) returned in step 6, use the blinding factor r cached in step 5 to perform a deblinding operation to obtain the authorized query information (t1, t2), where t1[i] = unmask(r -1 y1[i]), t2[i] = unmask(r -1 y2[i]), unmask(.) indicates unblinding, r -1 It is obtained by solving the modular inverse operation of the blinding factor r with respect to the order of the elliptic curve; Step 8: The user service program sends the authorized query information to the social media fake news detection provider for encrypted search, and counts the number of matching LSH keywords in the encrypted data set to derive the probability that the message is fake, and finally returns the query result to the user.

2. A privacy-preserving social media fake news detection method according to claim 1, characterized in that: The EC-OPRF function H(k,m)={0,1} l ×{0,1} * →{0,1} l By three hash functions H1:{0,1} * →{0,1} l ,H2:{0,1} * →E,H3:E→{0,1} l Implementation, that is, H(k,m)=H3(k·H2(H1(m))), where E is a point on the elliptic curve, H1 is the cryptographic hash function SHA256, and hash functions H2 and H3 are implemented by the elliptic curve cryptography library.

3. A privacy-preserving social media fake news detection method according to claim 1, characterized in that: The elliptic curve is the NIST P-192 elliptic curve.

4. The method for detecting fake social media messages with privacy protection according to claim 1, characterized in that: The locality sensitive hash (LSH) function is implemented using the Hamming-LSH algorithm.

5. The privacy-preserving social media fake news detection method according to claim 1, characterized in that: The mask(.) uses elliptic curve multiplication.

Images