Anti-fraud outbound call identification method for cross-number-segment voiceprint tracking

Abstract

The invention relates to an anti-fraud outbound recognition method for cross-number-segment voiceprint tracking, and the method comprises the steps: introducing a time domain voiceprint adversarial extraction mechanism, and adding an adversarial voice discriminator, an active inhibition device, a channel, and the interference of non-identity factors of language emotion in voiceprint modeling; constructing a speaker constant feature residual channel: taking stable syllable fragments including vowels and gutto vowels in call content as a reference extraction window, and filtering emotion or speech speed driving components; outputting a steady-state voiceprint contour flow as a unique basis for subsequent cross-number identity fusion; designing a dynamic number fusion graph based on the voiceprint steady-state flow; each time of voiceprint appearance is regarded as a single time point node, and all similar historical voiceprints form a serial number merging path; a number time sequence edge weight function is set; the occurrence time interval, the use frequency and the regional jump of the new number and the old number are synthesized, and whether the number belongs to the same user or not is evaluated; and establishing atlas connection between the voiceprint main body node and all number nodes thereof.

Description

Technical Field

[0001] The present invention relates to an anti-fraud outbound call identification method, in particular to an anti-fraud outbound call identification method based on cross-number segment voiceprint tracking. Background Art

[0002] The existing patent CN110070875A, an anti-telecom fraud method based on voice keyword detection and voiceprint, has to some extent solved the need to identify behavioral characteristics in early telecom fraud. In particular, it has certain practicality in responding to call behavior through keyword detection combined with a voiceprint ID scoring mechanism during a call. However, in the scenario of tracking complex, high-frequency cross-number fraud behavior, this technology has many structural and systematic deficiencies, and it is difficult to meet the current new trend of intelligent, concealed and coordinated fraud behavior.

[0003] From the perspective of the depth and stability of the identity tracking mechanism, it relies on sensitive keywords identified within a single call to trigger voiceprint extraction. The voiceprint ID is then compared with a database, a simple cumulative count is performed, and a threshold is used to determine whether to trigger an alert. This mechanism's logic chain is based on a linear path: keyword, voiceprint, and call count. The voiceprint itself serves only as a counting index and does not constitute a true identity anchor, nor does it introduce any structured attribution relationship with the number. When fraudsters change numbers or make calls across devices, even the slightest change in the voiceprint or the absence of a keyword trigger breaks the model's tracking, disrupting the identification chain. Regarding number diversity and structural merging capabilities, it only performs static analysis of numbers within a single call. There is no structural model to characterize the attribution evolution path between voiceprints and multiple numbers, nor is it able to determine whether a new number is part of the original number's evolution, let alone comprehensively assess the number's behavioral characteristics. When the same fraudster uses different numbers to call users in multiple regions within a single day, this technology can only treat each call as an independent event, losing the ability to structurally reveal the logic and consistency of their number-changing behavior.

[0004] From the perspective of the breadth and intelligence of behavioral feature recognition, the keyword detection mechanism used belongs to the traditional keyword library matching model, which essentially relies on a preset sensitive vocabulary. Its ability to identify non-keyword fraudulent tactics, variant words, or fraudulent behaviors expressed in obscure language is extremely low. At the same time, behavioral dimensions such as call frequency, time distribution, and call recipient type are completely untouched. Behavioral judgment is highly dependent on whether keywords are triggered, which is very prone to missed reports. In terms of the intelligence and early warning capabilities of the system response mechanism, it adopts a voiceprint trigger + keyword + count mode. It only issues a static prompt after the cumulative voiceprint ID exceeds the threshold. It is not real-time and cannot actively perceive the potential risks of unknown numbers. Its protection for users relies more on the accumulation of historical data and cannot achieve initial recognition and immediate blocking. From the perspective of system scalability and deployment complexity, although it has the characteristics of simple structure and convenient layout, its logical structure is destined to be only applicable to low-frequency, static, single-number, and multi-round triggered fraud behavior identification scenarios. It lacks the necessary modeling and response capabilities for the current complex behavior patterns of fraud gangs with quick strikes and withdrawals, wide number distribution, and cross-border collaboration, and it is difficult to adapt to the growing precision, intelligence, and structuring needs of the existing anti-fraud environment. Summary of the Invention

[0005] The purpose of the present invention is to provide an anti-fraud outbound call identification method for voiceprint tracking across number segments, thereby solving some of the drawbacks and shortcomings pointed out in the background technology.

[0006] The present invention solves the above-mentioned technical problems by adopting the following technical solution, which includes the following steps: A time-domain voiceprint adversarial extraction mechanism is introduced, and an adversarial voice change discriminator is added to voiceprint modeling to actively suppress interference from non-identity factors such as equipment, channels, and language emotions. A speaker-constant feature residual channel is constructed: stable syllable segments including vowels and glottal stops in the call content are used as the benchmark extraction window to filter out emotion- or speech-rate-driven components. A steady-state voiceprint contour stream is output as the sole basis for subsequent cross-number identity fusion. Design a dynamic number fusion graph based on the steady-state flow of voiceprints; each voiceprint appearance is regarded as a single time point node, and all similar historical voiceprints form a numbered merge path; Set the number time sequence edge weight function: comprehensively consider the time interval between the new number and the old number, the frequency of use, and the regional jump to evaluate whether the number belongs to the same user; and establish a graph connection between the voiceprint main node and all its number nodes.

[0007] Furthermore, the voiceprint subject node and number node fusion graph is expanded into a voiceprint, number, and behavior ternary graph; Extract the following information for each outbound call: Regularity of call duration in diurnal and night-time frequency patterns; The language content characteristics of keyword segments and speech rhythm; Profiles of vulnerable groups, such as the elderly and women; A clustering algorithm is used to discover highly consistent behavioral role clusters in the graph; these clusters are assigned labels such as fraud customer service role and investment mentor template.

[0008] Furthermore, the behavioral role clusters are formed into a role-risk mapping relationship; once a number in a voiceprint role cluster is reported, its risk will be propagated to other number nodes along the voiceprint anchoring path; when a number is suspected but not confirmed, its communication level will be temporarily lowered or a reminder label will be added, and it will not be immediately banned; all risk propagation paths will feed back to the voiceprint recognition module to improve the recognition of suspicious voiceprints in the next round of recognition.

[0009] Furthermore, the method of actively suppressing interference from non-identity factors such as equipment, channels, and language emotions includes: S1. Establish a voiceprint modeling mechanism based on the original time-domain speech signal to extract the speaker's voiceprint features from outbound voice data. In the voiceprint modeling process, introduce an adversarial voice changer discriminator to identify non-identity factors in the voice, including device type, call channel characteristics, and language emotion characteristics. S2. Simultaneously train the voiceprint extraction module and the adversarial voice change discriminator, using an adversarial optimization strategy to enable the voiceprint extraction module to actively shield non-identity factors distinguished by the discriminator; use the adversarially trained voiceprint representation to compare voices from different outbound numbers, so as to anchor and track the speaker identity across number segments; S3. Use voiceprint features as the basis for unified identity modeling, number merging, behavior analysis and risk assessment in the fraud outbound call identification system.

[0010] Furthermore, the voiceprint extraction module performs time-domain modeling on the original speech waveform without relying on spectrogram or Mel filter bank features; the adversarial voice changer discriminator is used to identify at least one of the following non-identity factors: speech channel characteristics, device recording characteristics, speech rate changes, emotional state and background noise distribution.

[0011] Furthermore, the adversarial training process includes: jointly training the voiceprint extraction module and the voice change discriminator, wherein the voiceprint extraction module aims to extract identity inseparable features, while the discriminator aims to distinguish channel / device features to the greatest extent, forming an adversarial structure; the voiceprint modeling process includes a time segment screening module based on the attention mechanism, which is used to identify and weaken non-stable speech segments containing excessive speech speed, explosive sounds or abnormal silence.

[0012] Furthermore, the method for establishing a graph connection between a voiceprint subject node and all its number nodes includes: A voiceprint subject node is established based on the speaker's voiceprint features; different phone numbers associated with the voiceprint subject are extracted from multiple outbound voice calls; and a number time-series edge weight function is set to calculate the association weight between the new number and the identified number. The function comprehensively considers at least the following factors: The time interval between the appearance of new and old numbers, the frequency of use of each number, and the geographical jump distance of each number; Based on the result of the edge weight function, determine whether the new number belongs to the voiceprint main node; connect the voiceprint main node with its corresponding multiple phone number nodes in a graph to form a one-to-many dynamic structural relationship; Through the newly established number time sequence edge weight function, it dynamically evaluates from multiple dimensions whether the newly appeared number should be attributed to the existing voiceprint subject, thereby achieving consistent tracking of identities across number segments. Voiceprint extraction is performed on each outbound voice data to generate a voiceprint vector , and aggregate a series of call records with similarity higher than the threshold to form the voiceprint main node . For all historical phone numbers associated with the voiceprint subject Store number and behavior data to form a number node set To determine the new number Whether it can be attributed to the voiceprint subject , introducing the following innovative edge weight function:

[0013] in,

[0014] Formula explanation: : Indicates the voiceprint subject With new number The attribution weight (ranging from 0 to 1, with the closer to 1 the more credible the attribution); : Comprehensive behavior influence function, which is a nonlinear combination function; : The time interval between the last call to the new and old numbers; : The average daily call frequency of the new number (in the past 72 hours); : Indicates the geographical jump coefficient, based on the cross-regional distance between the number registration location and the access location; : are the control coefficients of time, frequency and region (obtained by fitting historical data); exponential mapping function : Used to compress the results of multidimensional behavioral factors into attribution confidence between [0,1].

[0015] Graph connection logic: when ( is the attribution threshold, the recommended value is 0.75), the system will think that the new number is highly related to the voiceprint subject; Add as voiceprint subject node The subsidiary number node of point to The directed edge of , thereby updating the graph structure.

[0016] Furthermore, the graph structure is a dynamically updated structure, which adds, deletes or adjusts the weights of nodes and edges in real time according to the introduction of new call data; the calculation of regional jumps by the number time series edge weight function includes geographical similarity evaluation of the number registration location, call access point, and IP address information.

[0017] Furthermore, the time interval in the edge weight function is used to evaluate whether there is continuity in the calling behavior of two numbers. If the new number appears within a short period of time after the old number is terminated, it is considered to have a higher correlation. The graph structure supports the community detection algorithm, which identifies the existence of gang fraud by analyzing the number overlap or approximate structure between multiple voiceprint entities.

[0018] Furthermore, the graph connection structure can be combined with the behavior trajectory analysis module to further perform semantic clustering on number usage time periods, call frequencies and object features to assist in determining whether there are fraudulent behavior characteristics.

[0019] The anti-fraud outbound call identification method based on cross-number segment voiceprint tracking proposed in this invention has the following significant beneficial effects, and has outstanding advantages in terms of technological innovation, practical adaptability, and system intelligence: Using voiceprint features as cross-number identity anchors, starting from the speaking perspective, unified modeling of the speakers behind different numbers, different number segments, and even different operator accounts is achieved. Compared with traditional solutions that rely on number blacklists or static labels, this greatly improves the ability to track the frequent number changes of fraudsters, and structurally breaks the loophole of "changing numbers to evade responsibility." Through the innovative design of the number time-series edge weight function, it comprehensively considers the dimensions of call time interval, call frequency, and regional jumping behavior, dynamically calculates the probability of attribution between numbers, and constructs a scalable and evolvable voiceprint-number graph structure, supports incremental number merging and path visualization, and enhances the structured presentation of fraud trajectories.

[0020] In-depth analysis of structural overlap and call pattern similarities across multiple voiceprint entities effectively identifies collaborative fraud rings. It also combines script features, call target group characteristics, and graph structure modeling to construct highly interpretable fraud behavior profiles. Employing a modular graph engine and streaming voiceprint processing architecture, it boasts excellent engineering scalability and can be rapidly integrated into various deployment environments, including carrier networks, financial anti-fraud platforms, and cloud-based voice service platforms, meeting the real-time analysis and early warning needs of large-scale, high-frequency outbound voice call data. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a functional structure diagram of the cross-number identity modeling based on the time-domain voiceprint adversarial mechanism of the present invention.

[0022] Figure 2 This is the construction of the dynamic fusion graph of voiceprint anchor points across number segments and the risk propagation flow chart of the present invention.

[0023] Figure 3 This is the main flow chart of the cross-number voiceprint anchor point multi-number behavior graph merging and dynamic identification of the present invention.

[0024] Figure 4 This is a flowchart for implementing anti-fraud merging and risk diffusion based on the voiceprint-number-behavior ternary graph in a large-scale complaint scenario according to an embodiment of the present invention.

[0025] Figure 5 This is a flowchart of cross-number and cross-region multi-dimensional fusion anti-fraud dynamic graph modeling and affiliation and community identification in an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The following is a detailed description of the specific embodiments of the present invention with reference to the accompanying drawings.

[0027] Combined with attachment Figure 1 In an anti-fraud outbound call identification method that uses voiceprint tracking across multiple phone numbers, we propose an identity modeling method based on a time-domain voiceprint adversarial extraction mechanism to ensure stable identification of the same fraudster despite device switching, changing call environments, and deliberate voice changes. This method's core goal is to proactively remove non-identity factors from the original voice signal, retaining the speaker's physiologically stable portion as the core voiceprint representation, which serves as the anchor for cross-phone number identity fusion. The entire process, from data acquisition and preprocessing to model building and training strategies, is deeply designed and optimized around this goal.

[0028] During the data acquisition phase, the system collects real-time audio data from outbound calls from telecom operators and internet voice platforms (VoIP applications) at a uniform 16kHz sampling rate. The calls cover a variety of languages, accents, emotional expressions, and various device channels (mobile phones, headsets, and web microphones) to ensure the model's broad adaptability. The collected audio is aligned to the first 30 seconds of the call and annotated with metadata such as timestamp, channel source, number ID, and user feedback (whether or not it was a report) for subsequent training label assignment and model behavior interpretation.

[0029] The data preprocessing stage includes four steps: first, denoising the original speech (using Wiener filtering + adaptive spectral subtraction) to restore the original signal in the vocal tract; second, segmenting the speech into short time segments, each segment is 1 second with a 50% sliding window to increase the sample density; third, performing phoneme analysis on each segment to identify stable acoustic units, vowel segments (pronunciations such as a, e, and o), and glottal stops (plosives p, t, and k), which are usually least affected by emotions and speaking speed; fourth, marking these segments as stable syllable windows, and the rest as auxiliary segments, which are used to enhance the model's adversarial discrimination ability during training.

[0030] In terms of model structure, the entire voiceprint modeling consists of three components: a base extraction network, an adversarial discriminator, and a constant feature residual channel. The base extraction network utilizes a stack of 1D convolutions combined with a bidirectional GRU. Its goal is to directly extract speaker features from the speech signal in the time domain, avoiding unnecessary signal compression introduced by spectral transformations. This network outputs a 128-dimensional voiceprint vector. Two branches are introduced during training: the main task branch, which performs speaker identification (compared with the true identity label), and the adversarial task branch, consisting of a fully connected voice-changing discriminator with softmax, which identifies the device type, channel characteristics, speaking rate category, and emotional label of the voice. The adversarial mechanism uses the reverse gradient (Gradient Reversal Layer) technique to target the discriminator's recognition ability in the voiceprint extraction module. Specifically, the more the voiceprint features confuse the discriminator, the more purely it represents the identity.

[0031] At the same time, a speaker-invariant feature residual channel is introduced into the model's backbone. This channel specifically extracts information from the aforementioned stable syllable window and constructs an independent feature subspace to perform residual compensation before outputting the voiceprint through the main channel. This structural design aims to emphasize the contribution of the parts of the voice that are least affected by the device, channel, and emotion to identity recognition, while avoiding overall distortion caused by rapid speech and intonation deviations.

[0032] During training, a dual-task loss function is jointly optimized: on the one hand, the cross-entropy loss of speaker recognition is minimized to ensure that the extracted voiceprint can be stably mapped to a specific individual; on the other hand, the discriminant loss of the voice changer is maximized, that is, the voiceprint vector prevents the discriminator from accurately determining the channel or emotion category, thus strengthening its anti-interference ability. In addition, to enhance the model's generalization ability across device scenarios, a dynamic channel perturbation strategy is introduced. This strategy randomly changes the frequency band response, signal-to-noise ratio, and echo ratio of the call audio in the training set to simulate the audio degradation commonly seen in real-world fraud environments.

[0033] The final model output is a continuous stream of steady-state voiceprint contours. This means that for each call session, instead of extracting a single voiceprint vector, multiple nested representations of stable syllable windows are weighted and synthesized to form a dynamic voiceprint trajectory that can be compared across sessions. This trajectory serves as a unique identity anchor and is input into the subsequent voiceprint-number graph fusion module, used to uniformly identify the same speaker who appears repeatedly under multiple different phone numbers. This supports subsequent cross-number merging, gang behavior identification, and risk warning.

[0034] Combined with attachment Figure 2 To address the problem of fraudsters frequently changing phone numbers to evade tracking, we propose a method for constructing a dynamic number fusion graph using the speaker's voiceprint as an anchor. The core idea is to treat each recognized voiceprint steady-state stream as a single time node, numbering and merging the changing trajectories of these voiceprints on the timeline. This establishes a dynamic identity mapping relationship between multiple numbers, achieving identity unification and risk aggregation for the same voiceprint subject at different times and in different number segments. The entire system covers multiple technical aspects, including data acquisition, voiceprint extraction, graph structure generation, and evolution path tracking. The following is a detailed technical process and model implementation details.

[0035] During the data acquisition phase, the system extracts outbound voice data in real time from multiple voice sources, including telecom operator call logs, enterprise voice platforms, and VoIP social voice applications. This data is accompanied by metadata such as call timestamp, phone number, device identifier, call location IP, and user feedback tags (whether the call was disconnected or whether a complaint was filed). The sampling rate is standardized at 16kHz, and calls must contain at least three seconds of valid voice for further processing. To address artifacts, device variations, and cross-network compression noise, the system incorporates a voiceprint quality scoring mechanism at the data acquisition layer to eliminate voice segments with low signal-to-noise ratios, ensuring the availability and consistency of training data and real-time comparison samples.

[0036] In the preprocessing stage, all speech is first standardized, including waveform normalization, silence segment removal, echo cancellation, and energy balance adjustment; then the speech is sliced ​​in a 1-second window + 50% overlapping sliding manner to ensure that a voiceprint snapshot can be extracted every 0.5 seconds during the call; then, based on the trained time-domain voiceprint extraction model (composed of 1D convolution and bidirectional GRU stacking), the voiceprint vector is extracted from each speech window, and the voiceprint steady-state stream is extracted through the stable syllable recognition module (for locating vowel and glottal areas), that is, a multi-segment stable voiceprint vector sequence in a one-time conversation is formed, and averaged and aggregated to form a steady-state representation of a call.

[0037] Each extracted steady-state voiceprint stream is defined as a time node entity in the system. This node records the current voiceprint ID (which can be generated by clustering or by comparing with an existing database), the call timestamp, the call number, and its meta-attributes. The system establishes a time series channel for the voiceprint subject, namely the voiceprint trajectory. If a new voiceprint node exceeds the set threshold (above 0.88) after similarity comparison, it is judged to be a time-continuing event of the same subject and is connected to the voiceprint trajectory. To identify cross-number identity usage behavior, the system records the phone number used for each call in the trajectory and annotates any changes.

[0038] The system extracts all the number nodes from the aforementioned voiceprint trajectory and establishes merge edges with the main voiceprint node, forming a dynamic number fusion graph. The center of the graph is the voiceprint subject node, and the periphery is all the phone numbers that have appeared in it. Each edge represents the identity mapping relationship between a number and the voiceprint. This structure is not statically constructed, but rather a time-driven graph evolution model. That is, each time a new voiceprint steady-state flow appears, the system first determines its similarity and time series rationality with any existing voiceprint trajectory. Then, if the attribution relationship is confirmed, the system incorporates the node into the existing trajectory and connects the new number it carries to the dynamic graph. If the attribution cannot be confirmed, the system considers it to be a new voiceprint subject and creates a new trajectory and a new fusion graph separately.

[0039] In order to avoid different voiceprint entities being mistakenly merged due to noise, the system sets multiple auxiliary constraints when constructing the fusion graph: one is the time sliding window constraint, that is, only new numbers that appear within a certain time range (within 24 hours) are allowed to participate in the fusion judgment; the second is the rationality judgment of number jump, whether the new number is in the adjacent number segment with the old number, whether it has a similar location, and whether there are traces of device reuse; the third is the comparison of behavior patterns, if the outbound calling behavior of the new number (call frequency, call target portrait, call time) is highly consistent with the historical pattern of the old number, the merging weight is increased; the fourth is voiceprint quality verification, that is, before each merger, the voiceprint stability scoring mechanism must be used to confirm that the quality of the new node is not lower than the set threshold.

[0040] After the fusion graph is constructed, a multi-number evolution network structure with voiceprint identity as the core is formed. Each graph is a star-shaped structure in space, with the voiceprint anchor point in the center and multiple telephone number nodes on the periphery. In time, it is a linear evolution structure, marking the time point when each number first appeared. This graph is not only used for number merging, but also supports gang behavior identification: the system can identify different numbers that appear repeatedly in multiple graphs through number co-occurrence analysis between multiple voiceprint main graphs, thereby revealing the gang characteristics of number rotation and sharing. In addition, the graph also supports the cascade transmission of fraud risks: once a number is judged to be high-risk, the system will propagate the risk label to the central voiceprint node according to the graph structure, and simultaneously issue warnings to other numbers in the graph, so as to actively intercept numbers that have not yet been exposed.

[0041] Combined with attachment Figure 3 We propose a graph modeling method that uses voiceprint identity as an anchor and phone number behavior as a path. By assigning a behavior-aware time-series edge weight function to a number, we determine whether a new number belongs to an identified voiceprint subject. This function then constructs a multilateral graph relationship between voiceprint and phone number, supporting subsequent behavioral chain analysis and fraud ring identification. The entire technical solution encompasses data collection, behavioral feature extraction, model design, and graph structure generation, forming a systematic, real-time, and evolving multi-number tracking mechanism. Regarding data acquisition, the system collects call audio data and its metadata from voice call platforms, carrier switching equipment, or enterprise outbound call systems. This includes call start and end times, caller ID, call location, terminal type, access network, and user behavior (whether the call was answered or marked as spam). All data collection adheres to privacy protection principles and is only used after desensitization within legal boundaries. Audio data is standardized in 16kHz, 16-bit PCM format to ensure consistent voiceprint processing and is linked to call behavior data via number ID.

[0042] Data preprocessing includes three steps: the first is voice cleaning, which filters out silent and low-quality clips, and performs echo suppression and environmental noise reduction at the same time; the second is call behavior structuring, where the system extracts the behavioral sequence characteristics of each number, including call frequency (total number of calls in the past 72 hours), average call duration, call time distribution (daytime / late night), and dialing area statistics; the third is number trajectory reconstruction, which is to sort all the numbers that have appeared in the history of each user in time, and record the activation time (first call time) and deactivation time (last call time) of the number to evaluate the usage continuity relationship between numbers. The core of the model structure is the calculation module of the temporal edge weight function, which is used to dynamically determine whether two numbers belong to the same voiceprint entity. The function integrates three key dimensions: time interval, usage frequency, and regional jump. Specifically, the time interval refers to the time difference between the first appearance of the new number and the last use of the old number. If the time difference is very short (within a few hours), it indicates a replacement behavior; the frequency of use reflects whether the number is the main number or a temporary number. If the new number is used frequently in the initial stage and the call pattern is similar to that of the old number, the credibility is increased; regional jump calculates the geographical distance through the number's place of origin, the real-time call location, and the access IP address information. If the regional mutation is unreasonable (jumping directly from Fujian to Xinjiang), the credibility of its ownership relationship will be reduced.

[0043] These features are standardized and input into the scoring function to calculate the attribution weight. The system sets an empirical threshold (0.75). When the attribution weight exceeds this threshold, the system considers that the new number belongs to the current voiceprint subject and connects it in the graph. The graph structure is constructed in a dynamic evolutionary manner. The system establishes a central node for each voiceprint subject and peripheral nodes for the attributable phone number. The nodes are connected by directed edges. The edge weight is the attribution score result mentioned above. The entire graph is continuously updated and expanded over time, supporting reverse tracking and visual analysis. During the model training process, in order to improve the accuracy of attribution judgment, the system constructs a semi-supervised learning mechanism, in which some attribution labels come from user reporting results, number cancellation records or real-name registration information as positive attribution samples, and reverse samples include data on voiceprint mismatches or abnormal attribution behaviors, which are used to supervise the adjustment of the edge weight function. Through iterative optimization, the model is made more adaptable to the real fraud number switching strategy. In terms of graph application, the system supports correlation analysis between multiple graphs. If a number appears repeatedly in multiple voiceprint graphs, or the edge weights between multiple graphs are significantly similar, it can be judged as number sharing or operation by a fraud gang, thereby realizing cross-number, cross-person, and cross-regional linkage identification and joint early warning. Therefore, this method breaks through the limitations of traditional static analysis based on voiceprint similarity or number blacklists, and realizes dynamic, behavior-driven cross-number segment voiceprint fusion and risk identification, which has significant technological advancement and application promotion value.

[0044] Example 1: Combined with attachment Figure 4 The anti-fraud center of a major telecom operator received numerous complaints regarding a series of outbound phone calls impersonating customer service representatives. User reports revealed that these calls frequently claimed lost packages, system upgrades, or account anomalies, prompting users to provide their ID numbers and verification codes. These calls were frequently made across multiple provinces. Notably, the number of numbers involved was numerous and lacked consistent number segmentation, making it impossible to identify them based solely on the number.

[0045] The system first retrieved all outbound call audio and metadata from the call data center over the past 72 hours. This totaled 112,847 call records from 68,312 different numbers, with 86,201 calls lasting at least 15 seconds. After pre-screening, 11,002 records containing suspected fraud keywords or complaint flags were extracted.

[0046] During the voiceprint extraction phase, the system performs steady-state voiceprint extraction on each valid call. Using the ECAPA-TDNN architecture, it extracts a 128-dimensional speaker embedding vector from each call and identifies stable syllable windows. Clustering revealed 19 groups of voiceprint entities that recur across different calls, each associated with a large number of different phone number nodes. One such entity, S1, is associated with 26 different phone numbers (including carrier numbers such as 170, 171, and 199). This voiceprint entity is constructed as a central node in a graph, and its associated numbers form a dynamically expanding number fusion graph.

[0047] The system then enters the ternary graph construction phase. On the voiceprint-number graph centered on S1, the system further adds behavioral feature nodes for each call behavior, extracting the following three types of behavioral information from the call content and call behavior: First, we examine call time patterns. The system records each number's outgoing call timeline and calculates the frequency distribution of calls in four time periods: 0:00–6:00, 6:00–12:00, 12:00–18:00, and 18:00–24:00. For example, number N13 made 113 calls over two days, 92 of which occurred between 18:00 and 24:00, clearly indicating evening peak hour fraud activity. This pattern recurs across most numbers associated with this voiceprint subject.

[0048] Next, the system analyzes language content. Using the ASR module, the system transcribes all call speech into text, extracting key words and speech rhythm metrics. For example, call sample T2134 repeatedly contains high-risk phrases such as verification code, system update, and account closure. The system assigns a fraud keyword weight of 0.92. Furthermore, the system identifies the speech rhythm as high-speed with no pauses, typical of scripted speech. The language behavior feature vector is encoded as [high-speed, male, medium-frequency, keyword-intensive].

[0049] The third area involves caller profile identification. The system uses user information to match callees with demographic labels. Of these calls, 63% of the recipients were over 55 years old, and another 22% were non-smartphone users, placing them at medium to high risk for fraud. Of the 28 calls from number N7, 19 were to women over 60, creating a clear profile of those prone to fraud.

[0050] All behavioral features are encoded as behavior nodes in the graph structure and connected to the corresponding number nodes and voiceprint nodes via directed edges, forming a ternary graph structure of voiceprint-number-behavior. Edge weights in the graph use a multi-dimensional evaluation function based on TF-IDF and structural density to represent the typicality and consistency of behaviors.

[0051] Based on this ternary graph, the system applied a graph clustering algorithm to segment the entire graph into communities, identifying multiple highly consistent behavioral role clusters. Cluster C3 included voiceprint subject S1, 26 number nodes, and 712 behavioral nodes. Their call times were concentrated, their language structures were similar, and their victims were consistent, forming a highly similar fraudulent speech network.

[0052] Ultimately, the system assigned a label to each highly consistent cluster. Cluster C3 was flagged as a scam impersonating a courier customer service representative because its script included the phrase "Your package is lost, please enter the verification code to resend it." Cluster C6 was identified as an investment mentor scam, with its voice frequently referring to terms like return rate, futures, and market predictions, and its calls generally lasting long (most exceeding 300 seconds). As an empirical validation, the system randomly sampled 10 numbers in cluster C3 and conducted a call-back experiment. Seven of these numbers were shut down, while three, after the call was connected, still proactively requested users install the official platform app, which allows for remote control, posing as customer service representatives. This confirms that this cluster of individuals possesses strong fraudulent characteristics. Furthermore, when the newly identified number N87 was linked to a call, the system, with only one call record, classified it as cluster C3 based on its voiceprint anchor, time pattern (high frequency late at night), keyword structure (verification code + freeze + appeal), and victim profile (62-year-old male), achieving an immediate warning on the first outbound call.

[0053] To address the behavioral characteristics of fraudsters who frequently change numbers, after constructing a voiceprint-number-behavior ternary graph and forming highly consistent behavioral role clusters, we further designed and implemented a role-risk mapping mechanism based on these clusters. This mechanism propagates risk signals triggered by single-point reports throughout the entire graph structure, enabling dynamic risk prevention and control across multiple numbers and roles. First, each clustered behavioral role cluster (impersonating customer service, investment advisor) is uniquely numbered and mapped to a central voiceprint node to form a role cluster label library. Based on historical fraud precedents, keywords used in the conversation, and multi-dimensional risk factors derived from target group characteristics, the system assigns an initial risk level to each role cluster, which is then used as a dynamic attribute of the cluster. During operation, if the system detects that a number node in the role cluster has been reported by a user, or its calling behavior is marked as high-risk by the rule engine (short-term, high-frequency outbound calls combined with an unusual keyword combination), the risk propagation mechanism within the role cluster is immediately triggered. The risk propagation path follows the voiceprint anchor path, starting with the currently reported number and propagating back to the voiceprint subject node to which it is bound. From that voiceprint subject, the risk propagation spreads to all associated number nodes in the graph. The propagation weight is dynamically calculated based on the edge weight (voiceprint similarity × behavioral consistency). In character cluster C3 from the previous story, once number N87 is reported, the system immediately traces back to its voiceprint anchor point S1 and performs propagation calculations on the remaining numbers N1 to N26 bound to S1. If N12's call behavior within the past six hours is highly similar to N87's (keyword overlap > 0.85, and the call time period is the same), the system labels it a high-risk linked number. At the same time, for new numbers that haven't been reported but exhibit unusual behavior, such as N27, whose call behavior only partially overlaps or whose voiceprints have weak edge weights with anchor points, they are not immediately blocked. Instead, they are added to the risk observation list and, based on operational strategies, mitigated intervention measures are implemented. These include downgrading the number's communication level (reducing the daily outbound call limit), marking it as a suspicious number segment to alert users, and restricting its calls to high-risk groups (users over 65 years old). This ensures that legitimate users are not mistakenly blocked without definitive evidence. All paths, nodes, and dynamic edge weights involved in risk propagation are updated in real time within the graph structure and act as a reverse proxy for the voiceprint recognition model: call samples from reported numbers are fed back into the training set as counter-samples for the next round of voiceprint model training, improving the model's recognition sensitivity for the same character's speech and its adaptability to voice-changing attacks under different call conditions, forming a closed-loop learning mechanism of data, model, and graph.

[0054] A telecommunications anti-fraud center has received multiple user reports pointing to a fraud ring impersonating bank customer service representatives in outbound calls, informing victims of suspected money laundering and requiring them to cooperate with fund verification. The ring then tricked victims into downloading a remote control app and transferring funds. The caller numbers provided by the informants spanned multiple number ranges, including 170, 171, 198, and 199, and crossed carriers and provinces, involving at least 19 different numbers. Initially, the number segment analysis was impossible. The platform then fed the audio of these outbound calls into its voiceprint modeling and identity tracking system.

[0055] In step S1, the system first models the raw time-domain audio of these calls (sampling rate 16kHz, audio length 15-60 seconds). Instead of using traditional spectrograms or MFCC features, the system uses the original one-dimensional waveform as input, feeds it into a multi-layer convolutional network (Conv1D), and then connects it to a bidirectional GRU network to extract a 128-dimensional time-domain voiceprint feature vector. Simultaneously, the system activates the adversarial voice change discriminator module, which consists of a shallow, fully connected network. Its task is to determine the device type (Android phone, iPhone, headset, PC), the call channel (VoLTE, VoIP, PSTN), and whether the speaker is in a high-emotion state (calculating the emotion label based on speech rate, pitch fluctuation, and rate of change of voice amplitude). The model divides the audio into two paths: a main model and a discriminator. Using reverse gradient updating, the system constructs an adversarial architecture, aiming to trick the main voiceprint model into deceiving the discriminator, so that its output feature vectors are free of non-identifying attributes such as device, channel, or emotion, thereby preserving a purer speaker's physiological characteristics. In the training samples, the system introduced approximately 86,000 historical outbound call data, which contained voices of the same speaker recorded on multiple devices, and manually annotated the emotional level and channel source to ensure that the adversarial model can effectively learn what are identity-irrelevant factors.

[0056] In step S2, after completing initial training, the model extracted features from the audio of 19 calls from the reported incident. The system found that the voiceprint vectors extracted from 16 audio samples had a similarity greater than 0.92. Even though one of the calls involved a headset and another involved a regular Android phone on a 4G channel, and some of the voices contained obvious emotional suppression (with a tearful tone or a hurried tone), the voiceprint extraction module successfully stripped away these influencing factors and was able to stably identify them as the same subject. Based on this, the system assigned a unique voiceprint ID to this group of numbers, indicating successful identity anchoring. The model further retained three similar but below-threshold samples in the candidate pool, and as subsequent calls accumulated, it determined whether they should be incorporated into the voiceprint subject.

[0057] In step S3, the identified unified voiceprint feature vector is input into the system's number fusion and behavior graph module. The system categorizes 16 numbers under the voiceprint subject node and analyzes these numbers for behavioral labels. These findings include: call times concentrated between 2:00 PM and 5:00 PM on weekdays; a highly consistent speech rate (averaging 5.4 words per second); keywords such as account freezing, public security, and identity verification appearing significantly more frequently than in typical customer service calls; and the primary target audience for these calls is elderly women, accounting for 73%. The behavior graph assigns these nodes to role cluster C7, and based on the language and emotional induction structure, identifies this cluster as a fraudulent role impersonating a bank customer service representative.

[0058] During further empirical testing, the system received a user report regarding the number 199XXXX9132. This number had previously been assigned to the voiceprint anchor S_A by the model, and had high-weight connections to 15 other numbers in the graph. The system immediately propagated the risk label to the other number nodes in the graph according to the risk propagation mechanism, and added a risk reminder label to the unreported number 171XXXX4210, indicating that this number was suspected of making risky outbound calls and urged users to be cautious in their interactions before answering. At the same time, the call audio from the reported incident was fed back to the voiceprint training module as an annotated counter-sample, which was used in the next round of training to strengthen the model's resistance to voice changes, emotional interference, and device camouflage, forming a closed-loop self-learning mechanism.

[0059] To address the challenges of voiceprint feature extraction and masking, a voiceprint modeling method based on the original time-domain waveform is employed. Combined with an adversarial voice change discriminator, a voiceprint representation mechanism is constructed that proactively identifies and removes non-identifying factors, thereby improving the ability to consistently identify a speaker across devices, channels, and emotionally disturbed environments. During the preprocessing phase, the system does not utilize traditional spectrograms, Mel-filter banks (MFCCs), or other frequency-domain speech features. Instead, the model directly uses the original speech waveform in PCM format (sampling rate 16kHz, quantization precision 16bit) as the model input. This signal is segmented into multiple segments within a 0.5-second sliding window, each with a length of 8,000 points. The voiceprint extraction module utilizes a time-domain, one-dimensional convolutional neural network (Conv1D) architecture. The front-end consists of three stacked convolutional kernels to extract short-term pronunciation variation features. The back-end uses a bidirectional gated recurrent unit (Bi-GRU) to learn temporal context. The final output layer forms a 128-dimensional voiceprint vector embedding representation, which encodes the speaker's voice identity within the call segment. To ensure that the encoding does not carry non-identity factors, the system introduces a structurally independent adversarial voice changer discriminator module. This module inputs the same original audio data or its intermediate feature representation as the voiceprint extraction module, and sets five types of discrimination tasks: voice channel type recognition (VoIP / VoLTE / PSTN), device recording feature recognition (Android phone / PC headset / iPhone microphone), speech rate classification (fast / medium / slow), emotional state judgment (normal / excited / depressed) and background noise distribution estimation (such as silent / city / home / traffic scenes). Each task corresponds to a sub-discriminator branch, sharing the front-end feature extraction layer, and using a softmax classification structure to output the predicted labels of the five tasks. During the training process, the main voiceprint extraction model and the voice change discriminator are linked through the reverse gradient mechanism, and the entire model adopts an adversarial training method: the goal of the voiceprint extraction module is to retain identity information to the maximum extent while confusing the voice change discriminator, making it difficult to accurately identify the above non-identity information, while the discriminator is continuously optimized to improve its ability to distinguish channels, devices, speaking speed, emotions and background noise. The two reach a balance in the confrontation, so that the final extracted voiceprint embedding vector has stronger identity uniqueness and cross-scenario stability.In a scenario consistent with the aforementioned case, the system received call data from the fraudulent number 199XXXX9132. In the traditional voiceprint system, this number uses an Internet VoIP platform and makes outbound calls in a taxi. The voice contains slight traffic noise and suppressed emotions. If the spectral feature modeling method is used, the system will mistakenly judge that it is inconsistent with the previous S1 voiceprint subject due to the difference in energy distribution. The original waveform input actively strips away the channel compression features and noise mixing effects through the adversarial discriminator, so that the extracted voiceprint vector has a similarity of 0.918 with other numbers previously attributed to the S1 subject (170XXXX5123, 171XXXX0032), thereby achieving consistent anchoring across devices, scenarios, and channels, and incorporating it into the voiceprint subject fusion graph to participate in subsequent role identification and risk communication processes, verifying the ability to maintain identity stability and behavior tracking continuity under high-interference voice conditions.

[0060] The specific implementation process uses a voiceprint modeling method based on an adversarial training strategy combined with an attention time segment screening mechanism. The goal is to extract unique speaker characteristics from call audio that are unaffected by non-identity factors such as device differences, channel type, emotional state, and speech speed variations. At the same time, it eliminates unstable segments in the audio, ensuring consistent recognition across number segments and devices, even when fraudsters use different outbound calling methods, different numbers, and complex voice-changing disguises. The system received a real-time outbound audio segment from the platform from the number 170XXXX8123. This audio segment was sent via a network VoIP channel and contained several bursts of sound and areas of abnormal speech speed. Users reported this number for impersonating customer service. The system first inputs the raw speech waveform into the voiceprint extraction module. This module, composed of several 1D convolutional layers, bidirectional GRU units, and an attention mechanism layer, extracts physiological feature encodings from individual speech utterances in the time domain. During this process, the adversarial voice change discriminator module is integrated in parallel. This module shares some of the feature representations from the previous layers of the voiceprint extraction and is assigned a multi-channel recognition task, including device type classification (Android, Apple, headset), call path discrimination (PSTN, VoLTE, SIP), speech rate classification (slow, normal, fast), and background environment labeling (silence, urban noise, home echo). The model outputs the recognition probability of each interference source through a softmax function. During the joint training phase, the voiceprint extraction module aims to extract features that preserve identity invariance while deliberately suppressing any non-identity factors that can be recognized by the voice change discriminator. Specifically, a gradient reversal mechanism is used to reversely inject the loss of the voice change discriminator into the main voiceprint network, causing its optimization direction to be opposite to that of the discriminator, thus forming an adversarial structure. The specific loss function is composed of a speaker classification loss (the main task) and a multi-task loss (the adversarial task) for the voice change discriminator. Through joint optimization, the voiceprint representation's ability to withstand channel perturbations is continuously enhanced. To further prevent unstable speech segments from interfering with the model's discrimination results, an attention mechanism module is introduced at the end of the voiceprint extraction network to learn weights for the input speech sequence segments. This module assigns low weights to abnormal segments, areas of sudden speech rate changes, emotional outbursts, and areas of extreme silence based on the input segments' speech rate fluctuations, energy bursts, and silence ratio. After training, it can accurately identify segments with strong bursts but no identity features, and weaken them when features converge, thereby improving the stability and credibility of the final voiceprint embedding.In actual scenarios, the model successfully classified multiple numbers from different call conditions (fraudulent outbound calls using the 170, 171, and 199 number segments) into a unified voiceprint anchor S1. Even though some of the voices contained in-car noise, suppressed emotions, and non-standard microphone recordings, the method still achieved consistent recognition across call environments and successfully merged 16 fraudulent numbers into the subsequent risk behavior map, effectively supporting behavior clustering, risk diffusion, and early interception mechanisms, demonstrating the high feasibility and technical superiority of collaborative modeling of adversarial structures and attention modules in anti-fraud scenarios.

[0061] Example 2: Combined with attachment Figure 5 An anti-fraud monitoring platform received multiple user complaints about suspected fraudulent calls from financial customer service representatives. The callers, claiming account anomalies, tricked users into downloading an unknown app and provided multiple contact numbers for verification. Preliminary statistics revealed 32 numbers involved, spanning multiple carrier number blocks (170, 171, 198, and 199) and multiple regions. Traditional blacklist models are ineffective due to the large range of number blocks and short time intervals. The platform retrieved 1,238 relevant call records from the past three days and conducted voiceprint modeling and cross-number attribution analysis on this data.

[0062] The system first performs voiceprint extraction on each outbound voice call and generates a voiceprint vector using the ECAPA-TDNN network trained in time domain adversarial learning. In the preliminary comparison, the system found that although the numbers 170XXXX1234, 198XXXX8742 and 171XXXX7621 came from different regions, their voiceprint similarities were 0.935 and 0.942 respectively, which were highly consistent with the threshold of 0.92. Based on this, the system formed the voiceprint main node , which is regarded as the outbound calling behavior of the same person using different numbers at different times.

[0063] The system captured the new number 199XXXX4367 (marked as ) first appeared at 2:14 am for an outbound call, which lasted about 27 seconds. The voice was spoken quickly and in a strong tone. Belong to the voiceprint subject , the system analyzes its behavior variables and substitutes them into the newly set edge weight function: 1. Time interval :Record The first time it appeared was 2:14 am, The last call made by the attributed number 198XXXX8742 at 22:30 was about 3.73 hours apart. .

[0064] 2. Frequency of use : In the past 72 hours, this new number has made 21 outbound calls, an average of 7 calls per day. .

[0065] 3. Regional jump distance :The new number is registered in Langfang, Hebei, while the base station where the user accesses the call is in Nanjing, Jiangsu. The geographical straight-line distance is about 890 kilometers. .

[0066] The system sets the control coefficient as: : Used to improve the contribution of time proximity to attribution weight; : The frequency has a medium impact, due to the existence of template outbound calls for fraudulent activities; : Regional jumping has a strong exclusionary effect, but cross-provincial operations cannot be completely ruled out.

[0067] Substituting into the formula:

[0068] The results of the calculations are as follows: , the first item is

[0069] , the middle term is

[0070] , the last term is

[0071] but,

[0072] The attribution weight is calculated as:

[0073] Because the system sets the attribution weight threshold , this value is obviously too low, indicating that the number Should not be immediately attributed to The main reason is that its geographical jump is too large. In the absence of supporting equipment or social behavior evidence, the system maintains a cautious attribution strategy. This node is marked as a node to be observed and is not currently included in the fusion graph.

[0074] Compare this to another new number, 171XXXX7210, which appeared at 3:03 AM. Its registered location is Wuxi, Jiangsu, and its access point is Suzhou, 45 kilometers away. Its last call was to 170XXXX1234, with a time interval of only 1.1 hours. Its call frequency is 9 times per day. Substituting this into the formula: , ,

[0075]

[0076] , ,

[0077]

[0078]

[0079] The weight is much higher than the 0.75 threshold, and the system determines that the new number should belong to the voiceprint subject. , and create a line in the voiceprint map The directed edge pointing to 171XXXX7210 has an edge weight of 0.938, which updates the graph structure and triggers the role cluster update mechanism.

[0080] In order to solve the problem of traditional static recognition system failure caused by the dynamic characteristics of frequent number changes, channel switching, and cross-regional crimes in fraudulent behavior, a voiceprint-number fusion graph structure with dynamic update capability is introduced, and the regional jump factor in the key number time series edge weight function is deeply designed, so that it can accurately reflect the evolution trajectory of the fraud call network in real time during actual operation. After the system is deployed, when the platform receives new call data, the user makes or receives a call from the number 170XXXX7210, which is connected by Suzhou, Jiangsu. The system first extracts the voiceprint feature vector from the voice data The call is then compared with all existing voiceprint subject nodes in the current graph. If the similarity exceeds a set threshold (0.92), the call is assigned to that voiceprint subject. The system then uses the structure to determine whether the new number is an existing node in the graph. If so, the number's time-series edge weight function is used to evaluate the attribution and the graph structure is updated with the current behavior data.

[0081] The graph is represented by a directed weighted graph, with the voiceprint node at the center. , the periphery is multiple number nodes belonging to the subject , each edge Indicates the attribution association between the number and the voiceprint subject, and its weight is the result of the number time series edge weight function When new call data is introduced, the system will perform the following operations: If it is found that a number node has not been used for a long time (no call record within 30 days) and its edge weight is lower than the elimination threshold (0.3), the system will mark the node as a cooling node or delete it completely; if a number is frequently used recently and the new call significantly increases its edge weight with the voiceprint subject (from 0.65 to 0.89), the system will update the weight of the edge and re-order the associated edge set; if the newly added number node belongs to multiple voiceprint subject graphs (and is also associated with and All have edge weights > 0.75), the system will retain the optimal attribution through a conflict resolution mechanism (based on the voiceprint vector distance and call time distribution differences), and attach a weak binding label to the secondary subject for subsequent behavior graph clustering.

[0082] In the calculation of regional jumps, the system establishes three layers of location information dimensions for the regional attributes of each number: first, the number registration location, which is provided by the operator's location information, usually the provincial / municipal location identifier when the number was initially issued; second, the call access point location, which is obtained by the base station ID, Wi-Fi access point or GPS data during the call, with granularity down to the county level; third, the IP address resolution location, which is applicable to VoIP calls. The system uses the public network IP to check its geographic location, including the country, province, and city levels; the geographical distance is calculated for each of these three types of information (using the latitude and longitude comparison method), and different weights are assigned to synthesize the final regional jump distance. In actual operation, for example, the number N_{new} is registered in Tianjin, the access point is in Nanjing, and the IP address resolution shows it as Changping District, Beijing. The system calculates the distance from the registration location to the access point to be 935 kilometers, and the distance from the access point to the IP address is 104 kilometers. The system will use preset coefficients (such as a registration location weight of 0.4, an access point weight of 0.5, and an IP address weight of 0.1) to synthesize a final hop distance of approximately 693 kilometers, which is used to substitute into the edge weight function to calculate the final attribution confidence.

[0083] The dynamic graph structure is highly flexible and real-time when linked with user reports and operator blacklist data. When the number 198XXXX6321 was complained about, the system found that it was a node in the graph. , belonging to the voiceprint subject , then the risk label is reversely diffused to , and transmit the risk signal to other number nodes along the path with edge weight > 0.8 in the graph 、 、 , forming a local risk cluster, and adjusting the edge weight and node status in the graph, while triggering the behavior clustering module to The calling style, target object, and language pattern are re-clustered to decide whether to mark the entire graph as a high-risk role cluster.

[0084] In order to identify the identity continuity of fraudsters after frequent number changes and the behavioral characteristics of multi-voiceprint collaboration in gangs, an edge weight evaluation mechanism based on call time interval calculation is proposed, and a community detection algorithm on the graph structure is introduced to discover the structural overlap and collaborative behavior between multiple voiceprint subjects, thereby further supporting the traceability and association analysis of fraud gangs. The system first extracts the last active time of each number from the call data (that is, the timestamp of the last outbound call or call of the number), and records the first appearance time of each new number. The time difference between the two is defined as the time interval ΔT. This term in the edge weight function is designed to be , used to emphasize that the new number is likely to replace the old number in a short period of time; number 170XXXX8123 made the last outbound call, and the new number 171XXXX0049 appeared and dialed a similar target group. The time interval was 3.73 hours. The time weight item calculated by the function is ,in In practice, the value range is [2.0, 3.0] and can be dynamically set based on the timeliness of fraudulent behavior. This weighting item significantly improves the credibility of the attribution between numbers. When the time interval is less than 12 hours and the call behavior is highly similar, the system will assign a higher edge weight exceeding 0.85, thereby connecting the new number to the original voiceprint main node.

[0085] After the graph structure is updated, the system further executes the graph-level community detection algorithm to mine the group collaboration relationship between multiple voiceprint subjects. This algorithm is based on the graph embedding and clustering mechanism, and adopts the GEMSEC or Louvain algorithm. It takes the entire voiceprint-number graph as input to identify voiceprint communities with concentrated edge density and high structural repetition. In actual detection, the system analyzes the voiceprint graph containing 28 voiceprint nodes and 104 number nodes, and finds that the voiceprint subject 、 、 The connected numbers are highly overlapped (170XXXX9121 and 171XXXX6702 appear at the same time and In addition, the number switching paths are continuous in time, and the voice behaviors are highly similar (the speaking speed, keywords and target groups are all overlapped). The graph structure shows the characteristics of multi-center star-shaped overlap. After GEMSEC algorithm detection, the structure is divided into communities. Based on the shared numbers, unified call templates, and synchronized active time periods within the community, the system flagged it as a suspected coordinated fraud ring. Further comparisons of call logs revealed that multiple numbers within the community had initiated outbound calls to the same target user within a few minutes of each other, using different roles to induce them. One person impersonated a bank customer service representative, while another disguised themselves as a police officer for verification, confirming this as a typical multi-role coordinated fraud scheme. The system flagged the entire community as a high-risk structure and simultaneously pushed this structure and its evolutionary trends to the operator's system, supporting suspension and risk tracing.

[0086] In order to enhance the graph structure's ability to represent fraudulent behavior and further improve the semantic recognition accuracy of number attribution and behavior labels, a behavioral trajectory analysis module is introduced based on the dynamically constructed voiceprint-number fusion graph to extract the usage period, call frequency, and call object portrait features of each number node in the graph, and to achieve intelligent auxiliary judgment of suspicious outbound call behavior through the fusion of semantic clustering mechanism and graph structure. During system operation, for each number that has been attributed to the voiceprint main node, , the system extracts its behavioral trajectory information for 7 consecutive days from the call records, including but not limited to the number of calls per day, the distribution of call time periods, the average call duration, the type of location of the target number (whether it is a number segment commonly used by the elderly), and whether it is a first-time caller indicator. The system detected that the number 171XXXX8240 made an average of 27 outgoing calls per day in the past 7 days. The calls were concentrated in the 14:00-17:30 period, accounting for 78%. Among them, 61% of the number locations of the called users were highly overlapped with retirement communities and retiree registration areas. The average call duration was 46 seconds, and repeated dialing rarely occurred, indicating a strong tendency for first contact. Based on this information, the system generates a behavior vector [27, 0.78, 46, 0.61] for the number, and uniformly constructs the behavior vectors of all identified number nodes into a behavior matrix, which is input into the K-Means or DBSCAN semantic clustering module. The system selects the K-Means algorithm in this scenario and sets the initial number of clusters to 5. The clustering results show that the behavior vector falls into the behavior cluster The common characteristics of this cluster are high frequency of first-time outbound calls, focus on the elderly, and concentration in the afternoon. It is marked by the system's historical model as a typical fraud role impersonating government affairs or financial institutions. This cluster label will be reversely assigned to all related nodes in the graph structure, and the suspicious weight of the edges within the cluster will be increased through the edge weight adjustment mechanism. If a number has cross-region switching, voiceprint variation masking, and high-frequency calling behavior, this risk weight can be used to trigger the early warning model. After combining with user reporting information, the system confirms the behavior cluster. The 17 numbers in the survey were involved in at least 3 real fraud cases, and it was further discovered that these numbers belonged to the voiceprint subjects. 、 and , but their behavioral trajectories are highly consistent and their call patterns are similar in structure. The system immediately identifies the existence of shared scripts or collaborative crime among these three voiceprint subjects, and forms behavioral consistency edges across voiceprints in the graph, which is updated to a multi-layer graph structure and simultaneously pushed to the anti-fraud response center for manual analysis. Through the deep integration of behavioral trajectory analysis and graph structure, the system not only achieves risk quantification at the number level, but also establishes an interpretable fraud pattern recognition system through behavioral semantic clustering. It establishes a clear visual path and chain of evidence in complex fraud behaviors across number segments, regions, and identities, significantly improving the overall intelligence level and practical application capabilities of the recognition system.

Claims

1. Anti-fraud outbound call identification method based on voiceprint tracking across number segments, characterized by The following steps are involved: A time-domain voiceprint adversarial extraction mechanism is introduced, and an adversarial voice change discriminator is added to voiceprint modeling to actively suppress interference from non-identity factors such as equipment, channels, and language emotions. A speaker-constant feature residual channel is constructed: stable syllable segments including vowels and glottal stops in the call content are used as the benchmark extraction window to filter out emotion- or speech-rate-driven components. A steady-state voiceprint contour stream is output as the sole basis for subsequent cross-number identity fusion. Design a dynamic number fusion graph based on the steady-state flow of voiceprints; each voiceprint appearance is regarded as a single time point node, and all similar historical voiceprints form a numbered merge path; Set the number time sequence edge weight function: comprehensively consider the time interval between the new number and the old number, the frequency of use, and the regional jump to evaluate whether the number belongs to the same user; and establish a graph connection between the voiceprint main node and all its number nodes.

2. The anti-fraud outbound call identification method based on cross-number segment voiceprint tracking according to claim 1 is characterized in that The voiceprint subject node and number node fusion graph is expanded into a voiceprint, number, and behavior ternary graph; Extract the following information for each outbound call: Regularity of call duration in diurnal and night-time frequency patterns; The language content characteristics of keyword segments and speech rhythm; Profiles of vulnerable groups, such as the elderly and women; A clustering algorithm is used to discover highly consistent behavioral role clusters in the graph; these clusters are assigned labels such as fraud customer service role and investment mentor template.

3. The anti-fraud outbound call identification method based on cross-number segment voiceprint tracking according to claim 2 is characterized in that The behavioral role clusters are formed into a role-risk mapping relationship; once a number in a voiceprint role cluster is reported, its risk will be propagated to other number nodes along the voiceprint anchoring path; when a number is suspected but not confirmed, its communication level will be temporarily lowered or a reminder label will be added, and it will not be immediately banned; all risk propagation paths will feed back to the voiceprint recognition module to improve the recognition of suspicious voiceprints in the next round of recognition.

4. The anti-fraud outbound call identification method based on cross-number segment voiceprint tracking according to claim 3 is characterized in that The method for actively suppressing interference from equipment, channels, language, emotions and non-identity factors includes: S1. Establish a voiceprint modeling mechanism based on the original time-domain speech signal to extract the speaker's voiceprint features from outbound voice data. In the voiceprint modeling process, introduce an adversarial voice changer discriminator to identify non-identity factors in the voice, including device type, call channel characteristics, and language emotion characteristics. S2. Simultaneously train the voiceprint extraction module and the adversarial voice change discriminator, using an adversarial optimization strategy to enable the voiceprint extraction module to actively shield non-identity factors distinguished by the discriminator; use the adversarially trained voiceprint representation to compare voices from different outbound numbers, so as to anchor and track the speaker identity across number segments; S3. Use voiceprint features as the basis for unified identity modeling, number merging, behavior analysis and risk assessment in the fraud outbound call identification system.

5. The anti-fraud outbound call identification method based on cross-number segment voiceprint tracking according to claim 4 is characterized in that The voiceprint extraction module performs time-domain modeling on the original speech waveform without relying on spectrogram or Mel filter bank features; the adversarial voice changer discriminator is used to identify at least one of the following non-identity factors: speech channel characteristics, device recording characteristics, speech rate changes, emotional state, and background noise distribution.

6. The anti-fraud outbound call identification method of cross-number segment voiceprint tracking according to claim 5 is characterized in that The adversarial training process includes: jointly training a voiceprint extraction module and a voice change discriminator, wherein the voiceprint extraction module aims to extract identity inseparable features, while the discriminator aims to distinguish channel / device features to the greatest extent possible, forming an adversarial structure; the voiceprint modeling process includes a time segment screening module based on an attention mechanism, which is used to identify and weaken unstable speech segments containing excessive speech speed, explosive sounds or abnormal silence.

7. The anti-fraud outbound call identification method based on cross-number segment voiceprint tracking according to claim 1 is characterized in that The method for establishing a graph connection between a voiceprint subject node and all its number nodes includes: A voiceprint subject node is established based on the speaker's voiceprint features; different phone numbers associated with the voiceprint subject are extracted from multiple outbound voice calls; and a number time-series edge weight function is set to calculate the association weight between the new number and the identified number. The function comprehensively considers at least the following factors: The time interval between the appearance of new and old numbers, the frequency of use of each number, and the geographical jump distance of each number; Based on the result of the edge weight function, it is determined whether the new number belongs to the voiceprint main node; the voiceprint main node and its corresponding multiple telephone number nodes are connected in a graph to form a one-to-many dynamic structural relationship.

8. The anti-fraud outbound call identification method of cross-number segment voiceprint tracking according to claim 7 is characterized in that The graph structure is a dynamically updated structure, and the weights of nodes and edges are added, deleted or adjusted in real time according to the introduction of new call data; the calculation of regional jumps by the number time series edge weight function includes a geographical similarity assessment of the number registration location, call access point, and IP address information.

9. The anti-fraud outbound call identification method of cross-number segment voiceprint tracking according to claim 8 is characterized in that The time interval in the edge weight function is used to evaluate whether there is continuity in the calling behavior of two numbers. If the new number appears shortly after the old number is terminated, it is considered to have a higher correlation; The graph structure supports a community detection algorithm, which can identify existing group fraud by analyzing the number overlap or similar structure between multiple voiceprint entities.

10. The anti-fraud outbound call identification method based on cross-number segment voiceprint tracking according to claim 9 is characterized in that The graph connection structure can be combined with the behavior trajectory analysis module to further perform semantic clustering on number usage time periods, call frequencies, and object features to assist in determining whether there are fraudulent behavior characteristics.

Images