Anti-ai fraud method and electronic device

Abstract

Disclosed in the present invention are an anti-AI fraud method and an electronic device. The method comprises: when a target login account receives an interaction request, acquiring an account of a request user side; on the basis of a risky account identification rule, determining whether the account of the request user side is a risky account; if the account of the request user side is a risky account, on the basis of content indicated by the interaction request, using a preset risk identification rule to perform risk identification on said content, so as to determine whether said content is content generated on the basis of artificial intelligence (AI), wherein after risk identification is performed on content of different interaction types included in said content to obtain risk identification sub-results, a target risk identification result is determined on the basis of preset weight values; and on the basis of the target risk identification result, presenting preset prompt content in a user-side interface where the target login account is located. AI-based fraud is identified on the basis of AI, which can help users minimize the risk of fraud to the greatest extent.

Description

Method and electronic device for anti-AI fraud Technical Field

[0001] The present application relates to the field of information processing technology, and in particular to a method and electronic device for combating AI fraud. Background Art

[0002] With the advancement of technology and the rise of AI, the success rate of fraud using AI is higher, and there is currently no solution to this type of fraud.

[0003] Summary of the Invention

[0004] This application provides an anti-AI fraud method and electronic device to solve the problems existing in the related technology.

[0005] In a first aspect, the present invention provides a method for combating AI fraud, comprising obtaining the account of the requesting user terminal when a target login account receives an interaction request; determining whether the account of the requesting user terminal is a risk account based on a risk account identification rule; if it is a risk account, based on the content indicated by the interaction request, performing risk identification on the content using a preset risk identification rule to determine whether the content is content generated based on artificial intelligence, wherein, after performing risk identification on different types of interaction contained in the content to obtain sub-risk identification results, a target risk identification result is determined based on a preset weight value; based on the target risk identification result, a preset prompt content is presented in the user terminal interface where the target login account is located.

[0006] Optionally, risk identification is performed separately on the different types of interactions contained in the content to obtain sub-risk identification results, including: if the content in the interaction request is a picture, the content is identified based on the trained generative adversarial network; if the content in the interaction request is a video file, the content is identified based on the generative adversarial network to determine whether the video file is an artificial intelligence-generated video; if the content in the interaction request is a real-time video, whether it is a voice-controlled video is identified based on the trained network model; and / or, if the content in the interaction request is a real-time video containing a face, determining whether the face is a face-swapped face image based on a convolutional neural network.

[0007] Optionally, training a network model to identify whether it is a voice-controlled video includes: obtaining the voice-controlled video as a sample set, and using the sample set as input to train a recurrent neural network model, wherein the recurrent neural network model is h t =f(w x x t +w h h t-1 +b); where xt is the input vector, h t is the hidden state vector at time step t, w x is the input weight matrix, w h Hidden state weight matrix, b is the bias term, f is the activation function; the real-time video in the interaction request to be identified is input into the recurrent neural network model, and a first recognition result is output; the real-time video in the interaction request to be identified is input into the trained generative adversarial network, and a second recognition result is output; based on the preset recognition weights w1, w2, the two recognition methods obtain results, and the first final recognition result is calculated using a weighted sum method; based on the size relationship between the final recognition result and the first preset threshold, it is determined whether it is a voice-controlled video.

[0008] Optionally, determining whether the face is a face-swapped face image based on the trained convolutional neural network includes: obtaining the avatar data under the target login account, and extracting facial edge features based on the convolutional neural network; extracting pictures from real-time video frames containing faces, and performing edge extraction based on the convolutional neural network; determining the change in the edge feature vector of the face in each frame facial image; if the change is less than a first threshold, comparing the edge feature vector with the facial edge feature, wherein if the comparison result is less than a second threshold, it is a face-swapped face image.

[0009] Optionally, training a convolutional neural network to determine whether the face is a face-swapped face image includes: configuring a convolutional neural network, including: setting and determining a convolution kernel size and a Gaussian function standard deviation, wherein the Gaussian function standard deviation is σ=1, Use the gradient of the image Edge extraction is performed, where the gradient modulus It is used to reflect edge information. The image is composed of two pixels. x is defined as the value of the horizontal pixel and y is the value of the vertical pixel. The activation function is: z represents the output calculation of the neural network neuron Among them, the neurons of the neural network have n inputs, each input has a weight value and an input value; the i-th input is represented by w i Indicates its weight, with x i Represents the input value, b represents the bias, and the output value is the sum of all input weights multiplied by the input value plus the bias; obtain pictures or videos containing faces under the target login account as training samples, and train the convolutional neural network.

[0010] Optionally, if the content in the interaction request is a real-time video containing a face, determining whether the face is a face image after face-swapping based on a convolutional neural network includes: inputting the real-time video containing the face into the generative adversarial network, and outputting a third recognition result; inputting the real-time video containing the face into the convolutional neural network, and outputting a fourth recognition result; based on preset recognition weights w3, w4, two recognition methods obtain results, and a second final recognition result is calculated using a weighted summation method; based on the size relationship between the second final recognition result and the second preset threshold, determining whether it is a face image after face-swapping.

[0011] Optionally, the generative adversarial network is: Among them, G represents the generative model; D represents the discriminative model; IE represents the expected value; p data (x) indicates the original image or video, p z represents the image or video generated by training, V(D,G) represents the valuation function, x~p data (x) represents the random input real distribution sample data, where p data Represents the true distribution, the value of x is less than the total number of given true samples; z~P z (z) represents random input noise sample data, P z (z) noise data distribution, the value of z is less than the total number of given false samples generated; D(x) represents the result of distinguishing the real sample, G(z) in D(G(z)) represents the false data generated based on z in the generator, and D(G(z)) represents the result of the discriminator's judgment on the false data G(z). The above D(x) and D(G(z)) are the probabilities that the judgment result is true, and the value range is 0 to 1. In log(D(x)), because the log function is an increasing function, the larger the D(x) value, the larger the log(D(x)) value; in log(1-D(G(z))), the smaller the D(G(z)), the larger the function value. For the discriminator, the smaller the probability value of distinguishing the generated false data as true, the better, and the larger the overall expected value, the better; conversely, for the generator, the smaller the expected value, the better.

[0012] Optionally, train the discriminator using a gradient ascent strategy: Among them, x (i) Represents m real samples {x (1) , x (2) ,...x (m)};z (i) represents m noise samples {z (1) , z (2) ,...z (m)}; D(x (i)) represents the probability that the real sample is real; D(G(z (i) )) represents the noise sample z generated by G (i) The true probability of whether d Discriminator parameters, Represents the gradient.

[0013] Optionally, the determination of whether the account of the requesting user terminal is a risk account based on the risk account identification rules includes: judging whether the account of the requesting user terminal has changed the user terminal device, and if so, marking the account as a risk account; for an account marked as a risk account, judging whether the account of the requesting user terminal has preset usage behavior data, and if so, determining it as a risk account.

[0014] In a second aspect, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, an anti-AI fraud method provided in the first aspect is implemented.

[0015] The present invention discloses a method for combating AI fraud, comprising: when a target login account receives an interaction request, obtaining the account of the requesting user end; determining whether the account of the requesting user end is a risk account based on risk account identification rules; if it is a risk account, based on the content indicated by the interaction request, performing risk identification on the content using preset risk identification rules to determine whether the content is content generated based on artificial intelligence, wherein, after performing risk identification on the content of different interaction types contained in the content to obtain sub-risk identification results, the target risk identification result is determined based on a preset weight value; based on the target risk identification result, presenting preset prompt content in the user end interface where the target login account is located. AI-based fraud identification can help users minimize the risk of being defrauded. It overcomes the problem in related technologies that fraudulent behavior during the interaction process cannot be identified. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0017] FIG1 is a flow chart of an anti-AI fraud method of the present application;

[0018] FIG2 is a schematic diagram of the structure of a convolutional neural network provided in this application;

[0019] FIG3 is a schematic diagram of an electronic device provided by the present application corresponding to FIG1 . DETAILED DESCRIPTION

[0020] To make the purpose, technical solutions, and advantages of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this application and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0021] The following is an exemplary description of the anti-AI fraud method of the present application with reference to FIG1. ​​The method includes the following steps:

[0022] Step 101: When the target login account receives an interaction request, the account of the requesting user is obtained.

[0023] In this embodiment, the account may be an account of a social application, and online social interaction may be achieved by logging into the account.

[0024] Step 102: Determine whether the account of the requesting user terminal is a risky account based on risky account identification rules.

[0025] In this embodiment, a risky account refers to an account for which further determination is required as to whether the account user has a fraud risk.

[0026] As an optional implementation method of this embodiment, the determination of whether the account of the requesting user terminal is a risk account based on the risk account identification rules includes: judging whether the account of the requesting user terminal has changed the user terminal device, and if so, marking the account as a risk; for the account marked as a risk, judging whether the account of the requesting user terminal has preset usage behavior data, and if so, determining it as a risk account.

[0027] In this optional implementation, risky account labels are recorded when account users change devices (in the event of account theft or mobile phone loss, users will actively lock their accounts and take other actions to protect their accounts), and information such as IP location is recorded at the same time; for risky accounts, if there is any information related to transfers, investments, financial management, or money-sensitive information, the account will be marked with an increased risk level.

[0028] Step 103: If it is a risky account, based on the content indicated by the interaction request, the content is subjected to risk identification using preset risk identification rules to determine whether the content is content generated based on artificial intelligence. After risk identification is performed on the different types of interaction contained in the content to obtain sub-risk identification results, the target risk identification result is determined based on the preset weight value.

[0029] In this embodiment, if a risky account is determined, it is monitored to further determine whether the account poses a risk and provide recommendations. Different types of interactive content can include videos, images, and requested real-time video sent by the requesting user account. Predictions can be made for different types of content, and the final risk prediction result is determined based on the weights assigned to different types of interactive content.

[0030] Step 104: Based on the target risk identification result, a preset prompt content is presented in the user terminal interface where the target login account is located.

[0031] In this embodiment, if the risk identification result exceeds a preset threshold, a risk warning will be displayed on the user interface of the target login account. For users with a high risk level who are suspected of being theft, a warning will be given during information exchange, suggesting that they communicate in dialect or share their most profound experiences to verify the authenticity of the other party.

[0032] If it is not greater than the preset threshold, the monitoring will continue during the subsequent interaction process.

[0033] Common AI scams involve stealing a social media account, obtaining images, videos, and other data from the account, and then using AI to generate fraudulent images and videos, or using combined audio patterns to control video output or swap faces, convincing the victim. This embodiment uses AI to identify AI-generated images or videos, alerting potential victims to the possibility that the other party is using synthetic technology and fraud prevention measures. It also advises users to verify if the communication is real if they speak a common dialect or involve a shared experience of a specific, profound event. This effectively prevents AI scams.

[0034] As an optional implementation method of this embodiment, risk identification is performed on different types of interactive contents contained in the content to obtain sub-risk identification results, including: if the content in the interaction request is a picture, the content is identified based on the trained generative adversarial network; if the content in the interaction request is a video file, the content is identified based on the generative adversarial network to determine whether the video file is an artificial intelligence-generated video; if the content in the interaction request is a real-time video, whether it is a voice-controlled video is identified based on the trained network model; and / or, if the content in the interaction request is a real-time video containing a face, determining whether the face is a face-swapped face image based on a convolutional neural network.

[0035] In this optional implementation, user portraits and video data in the account can be collected and stored in advance, and the model of this application can be trained based on the collected data as samples. Different models are used for recognition based on the interactive content sent by the requesting user end.

[0036] Each piece of content is identified using a generative adversarial network designed in this embodiment, using its discriminator. Fraudsters use images and videos from stolen accounts and AI technology to generate images and videos. In this embodiment, after training, the generator attempts to generate realistic images or videos, while the discriminator attempts to determine whether these images or videos are realistic. Ultimately, the discriminator identifies the images and videos.

[0037] As an optional implementation of this embodiment, training a network model to identify whether it is a voice-controlled video includes: obtaining a voice-controlled video as a sample set, and using the sample set as input to train a recurrent neural network model, wherein the recurrent neural network model is h t =f(w x x t +w h h t-1 +b); where x t is the input vector, h t is the hidden state vector at time step t, w x is the input weight matrix, w h Hidden state weight matrix, b is the bias term, f is the activation function; the real-time video in the interaction request to be identified is input into the recurrent neural network model, and a first recognition result is output; the real-time video in the interaction request to be identified is input into the trained generative adversarial network, and a second recognition result is output; based on the set recognition weights w1, w2, the results obtained by the two recognition methods are finally calculated.

[0038] In this optional implementation, deep learning models, such as recurrent neural networks (RNNs), are used to classify or identify voice-activated videos. These models can learn and identify patterns and features of voice-activated videos based on training data.

[0039] RNN is represented by the following formula:

[0040] h t =f(w x x t +w h h t-1 +b); where: x t is the input vector, h t is the hidden state vector at time step t, w x is the input weight matrix, w hThe hidden state weight matrix, b, is the bias term, and f is the activation function. By continuously training with images, videos, and voice-activated video material, the RNN's memory capacity is increased, allowing it to identify whether a video is voice-activated. It's important to note that the subscripts x, h, and t in the above formulas for the recurrent neural network, representing the number of inputs, the number of hidden layers, and the time t, have no fixed range. The number of input and hidden layers in a neural network is analogous to N factors influencing a decision, and the number of these N factors varies in different scenarios. In mathematics, weights are generally represented by w (weight). The subscript x in w represents the input state weight, and the subscript h represents the hidden state weight. Using the same x and h for ease of understanding is crucial.

[0041] Furthermore, a generative adversarial network discriminator is used to identify AI voice-activated videos. After using RNN and GAN multimodal recognition, recognition weights w1 and w2 are set for the RNN and GAN, respectively. The final result of the two recognition methods is calculated as follows: the first final recognition result = w1 * first recognition result + w2 * second recognition result. If the first final recognition result is greater than or equal to a first preset threshold, it is a voice-activated video; if the first final recognition result is less than the first preset threshold, it is not a voice-activated video. w1 and w2 are the set weights, ranging from greater than 0 to less than 1.

[0042] As an optional implementation method of this embodiment, determining whether the face is a face-swapped face image based on a trained convolutional neural network includes: obtaining the avatar data under the target login account, and extracting facial edge features based on the convolutional neural network; extracting pictures from real-time video frames containing faces, and performing edge extraction based on the convolutional neural network; determining the change in the edge feature vector of the face in each frame facial image; if the change is less than a first threshold, comparing the edge feature vector with the facial edge feature, wherein if the comparison result is less than a second threshold, it is a face-swapped face image.

[0043] In this optional implementation, the avatar image data under the account is used as a training sample, and a convolutional neural network is used to extract facial feature vectors and avatar edge information. Real-time video frame images are obtained and facial edges are extracted using a deep learning convolutional neural network; the extracted facial edge feature vectors and color change values ​​of each frame are recorded. If the change in each facial edge feature vector is less than the set threshold, it is compared with the facial edge feature vector of the account avatar trained above. If it is less than the set threshold, it can be regarded as a face swap (in reality, the edge color of the avatar will change after moving in the environment, and the edge vector and color of the real-life video will be very different from the changes in expression and environment).

[0044] As an optional implementation of this embodiment, configuring a convolutional neural network includes: setting and determining the size of the convolution kernel and the standard deviation of the Gaussian function, wherein the standard deviation of the Gaussian function is σ=1, Use the gradient of the image Edge extraction is performed, where the gradient modulus Used to reflect edge information, where the activation function is z represents the output calculation of the neural network neuron Obtain pictures or videos containing human faces under the target login account as training samples to train the convolutional neural network.

[0045] In this embodiment, facial feature extraction is performed in the following steps: 1. Determine the size of the convolution kernel; 2. Set the standard deviation of the Gaussian function, for example, σ = 1. During the training process, the model continuously calculates the weight values ​​of each position of the convolution kernel and normalizes the weight values. The weight value w superscript i is continuously adjusted and changed through learning; 3. Edge extraction: using the gradient of the image Refers to the fastest direction of grayscale transformation, using the gradient modulus: The value reflects the edge information. Since the image is composed of two-dimensional pixels, x is defined as the value of the horizontal pixel point and y is the value of the vertical pixel point. The gradient change of the image is related to the partial derivative of the binary function.

[0046] Referring to the structure in Figure 2, conv is the convolution kernel; sigmoid is the activation function, which can be specifically expressed as z represents the output calculation of the neural network neuron pool is pooling; deconv is deconvolution; concat is channel superposition; m, n, minsize, etc. are used to set convolution parameters.

[0047] As an optional implementation method of this embodiment, if the content in the interaction request is a real-time video containing a face, determining whether the face is a face-swapped face image based on a convolutional neural network includes: inputting the real-time video containing the face into the generative adversarial network, and outputting a third recognition result; inputting the real-time video containing the face into the convolutional neural network, and outputting a fourth recognition result; based on the set recognition weights w3, w4, the results obtained by the two recognition methods are finally calculated.

[0048] In this optional implementation, for the recognition of AI face-changing, multimodal recognition such as convolutional neural network and generative adversarial network are used simultaneously, and recognition weights w3 and w4 are set for convolutional neural network and generative adversarial network respectively, and the final result is obtained after the two recognition methods obtain the results. When calculating, the second final recognition result = w3*third recognition result+w4*fourth recognition result. If the second final recognition result is greater than or equal to the second preset threshold, it is the face image after face-changing; if the first final recognition result is less than the second preset threshold, it is not the face image after face-changing. w3, w4 are the set weights, and the value range is greater than 0 and less than 1

[0049] As an optional implementation of this embodiment, the generative adversarial network is: Among them, G represents the generative model; D represents the discriminative model; IE represents the expected value; p data (x) indicates the original image or video, p z represents the image or video generated by training, V(D,G) represents the valuation function; x~p data (x) represents the random input real distribution sample data, where p data Represents the true distribution, the value of x is less than the total number of given true samples; z~P z (z) represents random input noise sample data, P z (z) noise data distribution, the value of z is less than the total number of given false samples generated; D(x) represents the result of discriminating real samples, G(z) in D(G(z)) represents the false data generated by the generator based on z, and D(G(z)) represents the result judged by the discriminator on the false data G(z). The above D(x) and D(G(z)) are the probabilities that the discrimination results are true, and the value range is 0.

[0050] In an alternative implementation, the generative adversarial network can be expressed as:

[0051] G represents the generative model; D represents the discriminative model; It means that we hope D(G(z)) will be as large as possible (1-D(G(z))) to achieve the minimum. In layman's terms, G hopes that the probability that D recognizes that the video image generated by G is true is as high as possible. It means that the higher the probability of D(x) identifying the original image, the better, and the lower the probability of D(G(z)) identifying the original image, the better. In layman's terms, D hopes to be able to greatly distinguish the authenticity of the image; IE represents the expected value; the present invention is to determine whether AI generates images and videos, and adopts The larger the better; data (x) represents the original image or video sample, p z Represents the image or video sample generated by training, and V(D,G) represents the valuation function.

[0052] As an optional implementation of this embodiment, the discriminator is trained using a gradient ascent strategy: Among them, x (i) Represents m real samples {x (1) , x (2) ,...x (m)};z (i) represents m noise samples {z (1) , z (2) ,...z (m)}; D(x (i) ) represents the probability that the real sample is real; D(G(z (i) )) represents the noise sample z generated by G (i) The true probability of whether d Discriminator parameters, Represents the gradient

[0053] In this embodiment, n represents the number of inputs in the input layer of the neural network, and has no fixed value range. The value range of i can be reflected in the formula as 1<=i<=n, and the value is [1,n].

[0054] In the expected value of the discriminator, among the m sample values ​​collected, the logarithm of the i-th real sample discriminant and the corresponding i-th generated sample are taken to obtain the discriminant of the i-th sample discriminator. Because the discriminator is not an elementary composite function, it is difficult to find the gradient by derivation or partial derivative, but it can be obtained by the derivation rule The principle is used to find the gradient. The discriminator discriminates the samples and uses the summation to obtain the total expected value of m samples, and then divides it by the total number of samples to obtain the discriminator gradient. Since the discriminator is used to identify fraud in this case, the discriminator objective function is sought to obtain the maximum value, and the gradient ascent method is used.

[0055] The present application also provides a computer-readable medium storing a computer program, which can be used to execute the method provided in FIG. 1 .

[0056] Figure 3 is a schematic diagram of the physical structure of an electronic device provided in an embodiment of the present invention. As shown in Figure 3, the electronic device 50 includes: a processor 501 (processor), a memory 502 (memory) and a bus 503; wherein, the processor 501 and the memory 502 communicate with each other through the bus 503; the processor 501 is used to call the program instructions in the memory 502 to execute the methods provided by the above-mentioned method embodiments.

[0057] This embodiment provides a non-transitory computer-readable storage medium, which stores computer instructions. The computer instructions enable a computer to execute the methods provided by the above method embodiments.

[0058] Those skilled in the art will understand that all or part of the steps of implementing the above-mentioned method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above-mentioned method embodiment; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk, etc. Various storage media that can store program codes.

[0059] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units. That is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0060] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, or of course, by means of hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or certain parts of the embodiment.

[0061] Although the present invention has been described in detail above using general descriptions and specific embodiments, it will be apparent to those skilled in the art that modifications and improvements may be made thereto. Therefore, such modifications and improvements, without departing from the spirit of the present invention, are intended to be within the scope of protection claimed herein.

Claims

1. A method for combating AI fraud, characterized in that: include: When the target login account receives an interaction request, it obtains the account of the requesting user; Determining whether the account of the requesting user terminal is a risky account based on risky account identification rules; If the account is at risk, based on the content indicated by the interaction request, the content is subjected to risk identification using preset risk identification rules to determine whether the content is AI-generated content. After risk identification is performed on each of the different types of interaction contained in the content to obtain sub-risk identification results, a target risk identification result is determined based on a preset weight value. Based on the target risk identification result, preset prompt content is presented in the user terminal interface where the target login account is located.

2. The anti-AI fraud method according to claim 1, characterized in that: Sub-risk identification results obtained by performing risk identification on the different types of interactions contained in the content include: If the content in the interaction request is a picture, performing a discrimination based on the trained generative adversarial network; If the content in the interaction request is a video file, the content is judged based on the generative adversarial network to determine whether the video file is an artificial intelligence-generated video; If the content in the interaction request is a real-time video, identifying whether it is a voice-controlled video based on the trained network model; and / or, If the content in the interaction request is a real-time video containing a face, it is determined based on a convolutional neural network whether the face is a face-swapped face image.

3. The anti-AI fraud method according to claim 2, characterized in that: Training a network model to identify whether a video is voice-activated involves: The voice-controlled video is obtained as a sample set, and the sample set is used as input to train the recurrent neural network model, wherein the recurrent neural network model is h t =f(w x x t +w h h t-1 +b); where x t is the input vector, h t is the hidden state vector at time step t, h t-1 It is h t The hidden state vector of the previous moment is used as the current h t The input of the state, w x is the input weight matrix, w h Hidden state weight matrix, b is the bias term, f is the activation function; Inputting the real-time video in the interaction request to be recognized into the recurrent neural network model, and outputting a first recognition result; Inputting the real-time video in the interaction request to be recognized into the trained generative adversarial network, and outputting a second recognition result; Based on the preset recognition weights w1, w2, the two recognition methods obtain results, and the first final recognition result is calculated using a weighted sum method; Based on the magnitude relationship between the final recognition result and a first preset threshold, it is determined whether the video is a voice-controlled video.

4. The anti-AI fraud method according to claim 2, characterized in that: Determining whether the face is a face-swapped face image based on the trained convolutional neural network includes: Obtaining the profile picture data of the target login account and extracting facial edge features based on a convolutional neural network; Extract images from real-time video frames containing faces and perform edge extraction based on convolutional neural networks; Determine the amount of change in the edge feature vector of the face in each frame of the facial image; If the change is less than a first threshold, the edge feature vector is compared with the facial edge feature, wherein if the comparison result is less than a second threshold, it is a face image after face replacement.

5. The anti-AI fraud method according to claim 4, characterized in that: Training a convolutional neural network to determine whether the face is a face-swapped face image includes: Configuring a convolutional neural network includes setting the size of the convolution kernel and the standard deviation of the Gaussian function, wherein the standard deviation of the Gaussian function is σ=1 and the standard distribution formula of the Gaussian function is Use the gradient of the image Edge extraction is performed, where the gradient modulus Used to reflect edge information, where the image is composed of two-dimensional pixels, x is defined as the value of the horizontal pixel point, y is defined as the value of the vertical pixel point, the gradient change of the image is related to the partial derivative of the binary function, the gradient formula α and β are the direction angles of the horizontal axis and the vertical axis, and the formula is abbreviated as Among them, the activation function is z represents the output calculation of the neural network neuron Among them, the neurons of the neural network have n inputs, each input has a weight value and an input value; the i-th input is represented by w i Indicates its weight, with x i Represents the input value, b represents the bias, and the output value is the sum of all input weights multiplied by the input value plus the bias; Obtain pictures or videos containing human faces under the target login account as training samples to train the convolutional neural network.

6. The anti-AI fraud method according to claim 2, characterized in that: If the content in the interaction request is a real-time video containing a human face, determining whether the human face is a face-swapped face image based on a convolutional neural network includes: Inputting a real-time video containing a human face into the generative adversarial network and outputting a third recognition result; Inputting a real-time video containing a human face into the convolutional neural network, and outputting a fourth recognition result; Based on the preset recognition weights w3, w4, the two recognition methods obtain results, and the second final recognition result is calculated using the weighted sum method; Based on the size relationship between the second final recognition result and the second preset threshold, it is determined whether it is a face image after face replacement.

7. The anti-AI fraud method according to any one of claims 2 to 6, characterized in that: The generative adversarial network is: Among them, G represents the generative model; D represents the discriminative model; IE represents the expected value; p data (x) indicates the original image or video, p z represents the image or video generated by training, V(D,G) represents the valuation function; x~p data (x) represents the random input real distribution sample data, where p data Represents the true distribution, the value of x is less than the total number of given true samples; z~P z (z) represents random input noise sample data, P z (z) noise data distribution, the value of z is less than the total number of given false samples generated; D(x) represents the result of discriminating the real sample, G(z) in D(G(z)) represents the false data generated by the generator based on z, and D(G(z)) represents the result judged by the discriminator on the false data G(z). The above D(x) and D(G(z)) are the probabilities that the discriminant result is true, ranging from 0 to 1. In log(D(x)), since the log function is an increasing function, the larger the D(x) value, the larger the log(D(x)) value; in log(1-D(G(z))), the smaller the D(G(z)), the larger the function value. For the discriminator, the smaller the probability of distinguishing false data as true, the better, and the larger the overall expected value, the better; conversely, for the generator, the smaller the expected value, the better.

8. The anti-AI fraud method according to claim 7, characterized in that: Use the gradient ascent strategy to train the discriminator: Among them, x (i) Represents m real samples {x (1) , x (2) ,...x (m) };z (i) represents m noise samples {z (1) , z (2) ,...z (m) }; D(x (i) ) represents the probability that the real sample is real; D(G(z (i) )) represents the noise sample z generated by G (i) The true probability of whether d Discriminator parameters, Represents the gradient.

9. The anti-AI fraud method according to claim 1, characterized in that: The determining whether the account of the requesting user terminal is a risky account based on the risky account identification rule includes: Determine whether the account of the requesting user terminal has changed the user terminal device, and if so, mark the account as risk; For an account that has been marked as risky, it is determined whether the account of the requesting user has preset usage behavior data. If so, it is determined to be a risky account.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method according to any one of claims 1 to 9 is implemented.

Images