A method, apparatus, and electronic device for detecting telecommunications fraud.

Abstract

This invention discloses a method, apparatus, and electronic device for detecting telecommunications fraud. The method includes: extracting features from a target audio sequence to obtain a sound embedding vector including emotional features; identifying the text of the target audio, comparing each word in the text with a preset lexicon across multiple language dimensions, and generating a language style vector; fusing the sound embedding vector and the language style vector using a deep learning model including convolutional neural networks and recurrent neural networks to obtain a fused feature vector including sound features and language style features; and detecting whether the target audio belongs to telecommunications fraud using the deep learning model. This invention has high accuracy and efficiency in detecting telecommunications fraud and can be widely applied in the field of telecommunications fraud detection.

Description

Technical Field

[0001] This invention relates to the field of telecommunications fraud detection technology, and in particular to a method, apparatus and electronic device for detecting telecommunications fraud. Background Technology

[0002] In recent years, the development of computing technology and security needs have increasingly highlighted the need for deception detection technology. Although automated deception detection is urgently needed in law enforcement, current technologies in this area remain quite limited. Existing technologies generally employ keyword or semantic analysis, which are effective against known types of telecommunications fraud, but are inefficient at detecting new and less obvious fraud, resulting in many new types of fraud going undetected and causing financial losses for the public.

[0003] Therefore, a telecommunications fraud detection scheme with high accuracy and efficiency is well worth studying. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a method, apparatus, and electronic device for detecting telecommunications fraud with high accuracy and efficiency.

[0005] One aspect of this invention provides a method for detecting telecommunications fraud, comprising:

[0006] Feature extraction is performed on the target audio using sound sequences to obtain sound embedding vectors that include emotional features;

[0007] The text of the target audio is identified, and each word in the text is compared with a preset vocabulary in multiple language dimensions to generate a language style vector.

[0008] By using a deep learning model that includes convolutional neural networks and recurrent neural networks, the sound embedding vector and the language style vector are fused to obtain a fused feature vector that includes sound features and language style features.

[0009] The deep learning model is used to detect whether the target audio belongs to telecommunications fraud.

[0010] Preferably, the step of extracting features from the target audio using a sound sequence to obtain a sound embedding vector including emotional features includes:

[0011] The target audio is divided into segments of 0.5 seconds each, with the segmentation labels remaining consistent with the original labels of the target audio. The audio is then resampled to a mono 16kHz.

[0012] The target audio is divided into frames using a 0.5s frame shift and a periodic Hanning window, and short-time Fourier transform and signal amplitude calculation are performed to obtain the spectrogram.

[0013] The spectrogram is mapped onto a Mel filter to obtain a Mel spectrum feature vector, which is used as the sound embedding vector.

[0014] Preferably, the step of comparing each word in the text with a preset lexicon across multiple language dimensions and generating a language style vector includes:

[0015] Each word in the text is compared with written language samples in a preset lexicon across multiple language dimensions to obtain a first comparison result;

[0016] Each word in the text is compared with spoken language samples in a preset lexicon across multiple language dimensions to obtain a second comparison result;

[0017] A language style vector is generated based on the first comparison result and the second comparison result.

[0018] Preferably, the training process of the deep learning model includes:

[0019] Multiple hidden layers are obtained by convolving the weight matrix with the lower-level activation matrix using a convolutional neural network and then using backpropagation to train the weights.

[0020] Dependencies are modeled in a time-series manner using recurrent neural networks, where each hidden layer state is a function of the previous state;

[0021] The spatial structure information in the training sound sequence is captured by a convolutional recurrent neural network, and the inherent spatial and temporal patterns in the training sound sequence are captured by a recurrent neural network.

[0022] Preferably, it further includes:

[0023] Web crawling technology was used to collect multiple voice messages from telecommunications networks from publicly available datasets;

[0024] The voice from the telecommunications network is input into the deep learning model to obtain the verification results;

[0025] If the verification result does not meet the set requirements, continue training the deep learning model until the verification result meets the set requirements.

[0026] Preferably, the step of fusing the sound embedding vector and the language style vector using a deep learning model including convolutional neural networks and recurrent neural networks to obtain a fused feature vector including sound features and language style features includes:

[0027] The sound embedding vector and the language style vector are input into a convolutional neural network and a recurrent neural network to obtain the first output of the convolutional neural network and the second output of the recurrent neural network.

[0028] The first output and the second output are fused to obtain a fused feature vector.

[0029] Another aspect of the present invention provides a device for detecting telecommunications fraud, comprising:

[0030] The emotion feature extraction unit is used to extract features from the target audio using the sound sequence to obtain a sound embedding vector that includes emotion features.

[0031] The language style feature extraction unit is used to identify the text of the target audio, compare each word in the text with a preset vocabulary in multiple language dimensions, and generate a language style vector.

[0032] The feature fusion unit is used to fuse the sound embedding vector and the language style vector through a deep learning model including convolutional neural networks and recurrent neural networks to obtain a fused feature vector including sound features and language style features.

[0033] The telecommunications fraud detection unit is used to detect whether the target audio belongs to telecommunications fraud through the deep learning model.

[0034] Another aspect of the present invention provides an electronic device, including a processor and a memory;

[0035] The memory is used to store programs;

[0036] The processor executes the program to implement the above-described method.

[0037] Another aspect of this invention provides a computer-readable storage medium storing a program that is executed by a processor to implement the above-described method.

[0038] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the method described above.

[0039] This invention extracts a sound embedding vector, including emotional features, from the target audio. It also compares each word in the identified target audio text with a pre-defined lexicon across multiple linguistic dimensions to fully understand the target audio's language style and generate a language style vector. Furthermore, it fuses the emotional features and language style features to obtain a fused feature vector. Finally, it uses a deep learning model, including convolutional neural networks and recurrent neural networks, to detect whether the target audio belongs to telecommunications fraud. This invention, by combining emotional and language style features, can identify subtle new types of telecommunications fraud, offering a wider and more accurate detection range compared to existing telecommunications fraud detection technologies. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A flowchart illustrating a method for detecting telecommunications fraud provided in an embodiment of the present invention;

[0042] Figure 2 An example flowchart of a method for detecting telecommunications fraud provided in an embodiment of the present invention;

[0043] Figure 3 A structural diagram of a deep learning model provided in an embodiment of the present invention;

[0044] Figure 4 Example code diagrams for a method of detecting telecommunications fraud provided in an embodiment of the present invention;

[0045] Figure 5 This is a structural block diagram of a telecommunications fraud detection device provided in an embodiment of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0047] Reference Figure 1 This invention provides a method for detecting telecommunications fraud, specifically including the following steps:

[0048] Step S100: Use the sound sequence to extract features from the target audio to obtain a sound embedding vector that includes emotional features.

[0049] Specifically, the process of extracting the sound embedding vector can include the following:

[0050] S1. Divide the target audio into segments of 0.5s in length, with the segmentation labels matching the original labels of the target audio, and resample to mono 16kHz.

[0051] S2. Using a frame shift of 0.5s and a periodic Hanning window, the target audio is divided into frames, and short-time Fourier transform and signal amplitude calculation are performed to obtain the spectrogram.

[0052] S3. Map the spectrogram onto a Mel filter to obtain a Mel spectrum feature vector, which is used as the sound embedding vector.

[0053] Step S110: Identify the text of the target audio, compare each word in the text with a preset vocabulary in multiple language dimensions, and generate a language style vector.

[0054] Specifically, the process of generating language style vectors may include the following:

[0055] S1. Compare each word in the text with written language samples in a preset lexicon across multiple language dimensions to obtain a first comparison result.

[0056] S2. Compare each word in the text with spoken language samples in a preset lexicon across multiple language dimensions to obtain a second comparison result.

[0057] S3. Generate a language style vector based on the first comparison result and the second comparison result.

[0058] In one optional implementation, the word library preset by the present invention can contain more than 2,000 words, divided into 72 language dimensions. When the number of words and language dimensions in the word library reach a certain scale, the generated language style vector can be very close to the real language style of the target audio.

[0059] Step S120: Using a deep learning model including convolutional neural networks and recurrent neural networks, the sound embedding vector and the language style vector are fused to obtain a fused feature vector including sound features and language style features.

[0060] Specifically, the training process of a deep learning model can be described as follows:

[0061] S1. The weight matrix and the lower-level activation matrix are convolved through a convolutional neural network. Backpropagation is used to train the weights to obtain multiple hidden layers.

[0062] S2. Dependencies are modeled in a time series manner using a recurrent neural network, where each hidden layer state is a function of the previous state.

[0063] S3. Capture the spatial structure information in the training sound sequence through a convolutional recurrent neural network, and capture the inherent spatial and temporal patterns in the training sound sequence through a recurrent neural network.

[0064] To verify the training effect of the deep learning model, this invention can further include a process for verifying the deep learning model, which may specifically include the following:

[0065] S1. Use web crawling technology to collect multiple voice messages from telecommunications networks from publicly available datasets.

[0066] S2. Input the voice from the telecommunications network into the deep learning model to obtain the verification results.

[0067] S3. If the verification result does not meet the set requirements, continue training the deep learning model until the verification result meets the set requirements.

[0068] Then, the sound embedding vector and language style vector can be input into a validated deep learning model and the features can be fused, specifically including the following:

[0069] S1. Input the sound embedding vector and the language style vector into a convolutional neural network and a recurrent neural network to obtain the first output of the convolutional neural network and the second output of the recurrent neural network.

[0070] S2. Merge the first output and the second output to obtain the fused feature vector.

[0071] Step S130: Detect whether the target audio belongs to telecommunications fraud audio using the deep learning model.

[0072] Specifically, based on deep learning models and fused feature vectors, it is possible to quickly and accurately identify whether the target audio belongs to telecommunications fraud. If so, the target audio can be marked in time and further processing can be carried out.

[0073] To describe the invention in more detail, specific examples will be used to illustrate its practical application.

[0074] Specifically, refer to Figure 2 This invention provides an example flowchart of a method for detecting telecommunications fraud. (Refer to...) Figure 3 This invention provides a structural diagram of a deep learning model. (Refer to...) Figure 4 The present invention provides an example code diagram for implementing a method for detecting telecommunications fraud.

[0075] Step 1: Data preparation and preprocessing.

[0076] This invention can obtain representative fraudulent and non-fraudulent voice data from telecommunications operators, and clean and preprocess the dataset using techniques such as tagging and noise reduction.

[0077] Step 2: Extract emotional features from the voice.

[0078] The audio obtained in the first step is used to extract features from the sound sequence to obtain the sound embedding vector, which is the sound feature corresponding to each sequence.

[0079] The specific feature extraction steps are as follows:

[0080] a. Divide all audio into 0.5s frame lengths, keeping the segmentation module labels consistent with the original audio labels, and resample to mono 16kHz.

[0081] b. A frame shift of 0.5s and a periodic Hanning window are used for framing, a short-time Fourier transform is performed, and then the signal amplitude is used to calculate the spectrogram.

[0082] c. Map the spectrogram obtained in step b to the Mel filter and calculate the Mel spectrum eigenvector.

[0083] Step 3: Extract language style features.

[0084] By creating a language archive, linguistic logic analysis and word counting are used—a text analysis program that analyzes written or spoken samples word by word. Each word is compared to a file of over 2000 words, divided into 72 linguistic dimensions, generating a language style vector.

[0085] Step 4: Feature fusion and model training.

[0086] This invention employs a Convolutional Recurrent Multi-Kernel Learning (CRMKL) model for training, utilizing a Convolutional Neural Network (CNN). In this CNN, sounds at times t and t+1 are combined into an image. In a deep CNN, each hidden layer is obtained by convolving the weight matrix with the activation matrix of the next lower layer and training the weights using backpropagation. A Recurrent Neural Network (RNN) with feedback connections between neurons can model dependencies in a time-series manner. Each hidden layer state is a function of the previous state and can be further extended to all previous states. The convolutional RNN captures spatial structural information in the sound sequence, allowing the model to capture the inherent spatial and temporal patterns within the sound sequence. By combining CNN and RNN models and fusing data features from sound and language, this approach offers improvements in both speed and accuracy compared to traditional MKL multi-kernel learning.

[0087] Step 5: Model validation.

[0088] To verify the effectiveness of using deep learning methods to detect web pages and online posts through web crawling technology, approximately 10,000 voice messages from telecommunications networks were collected, most of which were not fraudulent. The results obtained using the model showed an F1 score of 0.9, indicating that the deep learning model in this embodiment of the invention has excellent detection performance.

[0089] Reference Figure 5 This invention provides a device for detecting telecommunications fraud, comprising:

[0090] The emotion feature extraction unit is used to extract features from the target audio using the sound sequence to obtain a sound embedding vector that includes emotion features.

[0091] The language style feature extraction unit is used to identify the text of the target audio, compare each word in the text with a preset vocabulary in multiple language dimensions, and generate a language style vector.

[0092] The feature fusion unit is used to fuse the sound embedding vector and the language style vector through a deep learning model including convolutional neural networks and recurrent neural networks to obtain a fused feature vector including sound features and language style features.

[0093] The telecommunications fraud detection unit is used to detect whether the target audio belongs to telecommunications fraud through the deep learning model.

[0094] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to perform... Figure 1 The method shown.

[0095] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is altered and sub-operations described as part of a larger operation are executed independently.

[0096] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0097] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0098] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0099] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0100] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0101] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0102] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0103] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A method of detecting telecommunications fraud, characterized by, The method comprises: performing feature extraction on a target audio using a sound sequence to obtain a sound embedding vector comprising emotional features; identifying text of the target audio, comparing each word in the text with a preset vocabulary in multiple language dimensions, and generating a language style vector; the words in the preset vocabulary exceed 2000 and are divided into 72 language dimensions; fusing the sound embedding vector and the language style vector through a deep learning model comprising a convolutional neural network and a recurrent neural network to obtain a fused feature vector comprising sound features and language style features; detecting whether the target audio belongs to a telecom fraud audio through the deep learning model; the comparing each word in the text with the preset vocabulary in multiple language dimensions and generating a language style vector comprises: comparing each word in the text with written language samples in the preset vocabulary in multiple language dimensions to obtain a first comparison result; comparing each word in the text with spoken language samples in the preset vocabulary in multiple language dimensions to obtain a second comparison result; generating a language style vector according to the first comparison result and the second comparison result; the training process of the deep learning model comprises: training weights through a convolutional neural network by convolving a weight matrix with an activation matrix of a lower layer to obtain multiple hidden layers using backpropagation; modeling the dependency relationship in a time sequence through a recurrent neural network, and each hidden layer state is a function of the previous state; capturing spatial structure information in the training sound sequence through a convolutional recurrent neural network, and capturing inherent spatial and temporal patterns in the training sound sequence through a recurrent neural network.

2. The method of claim 1, wherein the method further comprises: The method for performing feature extraction on a target audio using a sound sequence to obtain a sound embedding vector comprising emotional features comprises: dividing the target audio into segments with a time length of 0.5s, and the division label is consistent with the original label of the target audio, and resampling into a single channel at 16kHZ; performing framing on the divided target audio using a frame shift of 0.5s and a periodic Hann window, and performing short-time Fourier transform and signal amplitude calculation to obtain a spectrogram; mapping the spectrogram to a mel filter to obtain a mel spectrum feature vector as a sound embedding vector.

3. The method of claim 1, wherein the method further comprises: The method further comprises: collecting multiple telecom network voices using a web crawler technique on a public data set; inputting the telecom network voices into the deep learning model to obtain a verification result; if the verification result does not meet the set requirement, continuing to train the deep learning model until the verification result meets the set requirement.

4. The method of claim 1, wherein, The method for fusing the sound embedding vector and the language style vector through a deep learning model comprising a convolutional neural network and a recurrent neural network to obtain a fused feature vector comprising sound features and language style features comprises: inputting the sound embedding vector and the language style vector into the convolutional neural network and the recurrent neural network to obtain a first output of the convolutional neural network and a second output of the recurrent neural network; fusing the first output and the second output to obtain a fused feature vector.

5. A device for detecting telecommunication fraud, characterized in that The method comprises: The emotion feature extraction unit is configured to perform feature extraction on the sound sequence of the target audio to obtain a sound embedding vector including emotion features. The language style feature extraction unit is configured to identify the text of the target audio, compare each word in the text with a preset word bank in multiple language dimensions, and generate a language style vector; the words in the preset word bank exceed 2000 and are divided into 72 language dimensions. The feature fusion unit is configured to fuse the sound embedding vector and the language style vector by a deep learning model including a convolutional neural network and a recurrent neural network to obtain a fusion feature vector including sound features and language style features. The telecommunications fraud detection unit is configured to detect whether the target audio belongs to a telecommunications fraud audio by the deep learning model. The comparison of each word in the text with the preset word bank in multiple language dimensions and the generation of the language style vector include: The comparison of each word in the text with written language samples in the preset word bank in multiple language dimensions to obtain a first comparison result; The comparison of each word in the text with spoken language samples in the preset word bank in multiple language dimensions to obtain a second comparison result; The generation of the language style vector according to the first comparison result and the second comparison result; The training process of the deep learning model includes: The use of back propagation to train the weight by convolution of the weight matrix and the lower layer activation matrix by the convolutional neural network to obtain multiple hidden layers; The modeling of the dependency relationship in time sequence by the recurrent neural network, and the state of each hidden layer being a function of the previous state; The capture of the spatial structure information in the training sound sequence by the convolutional recurrent neural network, and the capture of the inherent spatial and temporal patterns in the training sound sequence by the recurrent neural network. The processor and the memory; 6. An electronic device, comprising: The memory is configured to store a program; The processor executes the program to implement the method of any one of claims 1 to 4. The storage medium stores a program, and the program is executed by the processor to implement the method of any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer program is executed by the processor to implement the method of any one of claims 1 to 4.

8. A computer program product comprising a computer program, characterized in that, ​

Images