A visual detection anti-cheating method for a chess interaction device

Abstract

The application discloses a visual detection anti-cheating method of a chess interaction device, which comprises the following steps: acquiring a face image of a testee, calculating a sight gazing area, and mapping the sight gazing area to a chessboard to obtain a gazing area sequence; simultaneously detecting a chess piece dropping event to obtain a chess piece dropping coordinate; extracting a gazing sequence in a time window before the chess piece dropping; calculating a sight-chess piece dropping matching coefficient according to a matching ratio of the gazing area and the chess piece dropping area; correcting an original AI coincidence degree calculated by an AI engine by using the coefficient; generating a behavior attribution coincidence degree; and determining a cheating risk according to the behavior attribution coincidence degree. The method introduces real-time sight gazing behavior into AI coincidence degree evaluation, establishes a causal attribution mechanism of visual attention and chess piece dropping decision, can distinguish between normal high coincidence degree of autonomous thinking and abnormal high coincidence degree that cannot be explained by sight, and solves the problem that high-level players are prone to false positives due to the disconnection between visual monitoring and AI coincidence degree analysis.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and intelligent monitoring technology, and in particular to a visual detection method for preventing cheating in chess-like interactive devices. Background Technology

[0002] The widespread adoption of online Go grading exams and tournaments has made ensuring fairness in remote play a critical technical challenge. Unlike in-person exams where invigilators can directly observe candidates, online exams place candidates in private spaces. A single camera set up by the candidate themselves is insufficient to effectively monitor their visual attention, significantly increasing the risk of cheating.

[0003] To address the aforementioned issues, several anti-cheating solutions have been proposed in existing technologies. One type of solution relies on behavioral monitoring of invigilation videos. This involves analyzing video images of the candidate's area using facial recognition, human detection, and object detection to identify whether multiple people are present in the frame, whether candidates have left their seats, and whether prohibited items such as mobile phones or Bluetooth headsets are present. While this type of solution can detect some obvious violations, its detection logic essentially identifies isolated behavioral events or static objects. It cannot determine whether the candidate's visual attention is truly focused on the chessboard, nor can it distinguish between the two behavioral patterns: "making a move after independent observation and thought" and "making a move with external assistance without observing the chessboard." When a candidate only glances at external cues briefly before making a move, maintaining focus on the chessboard for the rest of the time, simple behavioral detection is insufficient for effective identification.

[0004] Another approach relies on AI consistency analysis. This involves comparing a player's game record with the recommended moves of a Go AI engine, calculating the degree of consistency between each move and the AI's first choice, and using the overall consistency score as the basis for determining cheating. A higher consistency score indicates that the player's move selection is closer to the AI's judgment, increasing the suspicion of cheating. However, this approach has an inherent flaw: high-level players may also make moves that closely match the AI's during normal gameplay. If cheating is judged solely based on the consistency score, it's easy to misjudge normal high-level play as cheating, leading to false alarms. Conversely, cheaters can lower the overall consistency score by alternating between high-consistency and low-consistency moves to evade detection. The root cause of this flaw lies in the fact that existing AI consistency analysis schemes only focus on "whether the move result is consistent with the AI," without considering "the visual attention state under which the move result was made," lacking a mechanism to incorporate visual behavioral information into the consistency evaluation system.

[0005] In summary, current technologies for visual behavior monitoring and AI consistency analysis operate independently. Visual detection results are presented only as discrete alarm labels, while AI consistency is presented only as isolated statistical values, without establishing a causal relationship between the two. Existing technologies do not yet provide a method for structured analysis of a player's real-time visual attention behavior during gameplay, directly influencing the correction of AI consistency values. This would allow the consistency index to carry behavioral attribution information, distinguishing between "high consistency resulting from autonomous observation and thought" and "high consistency that cannot be explained by visual attention behavior." This deficiency constitutes a critical technical problem that urgently needs to be addressed in the field of visual detection and anti-cheating technology for chess-like interactive devices. Summary of the Invention

[0006] The purpose of this invention is to provide a visual detection method for preventing cheating in chess interactive devices. By incorporating the examinee's real-time gaze behavior into an AI consistency evaluation system, a causal attribution mechanism between visual attention and move-making decisions is established. This enables the consistency index to distinguish between normal high consistency generated by autonomous observation and thinking and abnormally high consistency that cannot be explained by gaze behavior. This significantly improves the accuracy of online grading cheating detection and solves the problem in existing technologies where visual behavior monitoring and AI consistency analysis are disconnected, leading to high-level players being easily falsely reported and concealed cheating behaviors being difficult to effectively identify.

[0007] To address the aforementioned technical problems, a first aspect of this invention provides a visual detection anti-cheating method for a chess-playing interactive device, comprising the following steps: S1: The candidate's facial image is acquired by the first image acquisition module pointing towards the candidate's face, the candidate's gaze area is calculated in real time, and the gaze area is mapped to a pre-calibrated chessboard coordinate system to obtain a gaze chessboard region sequence; wherein, the gaze chessboard region is the region in which the mapping point of the gaze area falls on the chessboard plane after the chessboard plane is divided into multiple preset regions. S2: Detect the move event by using the second image acquisition module pointing to the chessboard area, obtain the coordinates of the current move in the chessboard coordinate system, and record the move time; S3: For the current move, extract the gaze chessboard region sequence within a preset time window before the move. Calculate the gaze-move matching coefficient based on the proportion of valid gaze frames matching the region where the move is located to the total number of valid gaze frames within the preset time window. When the proportion of valid gaze frames within the preset time window is lower than a preset threshold, mark the gaze-move matching coefficient as invalid. S4: Obtain the original AI consistency calculated by the AI ​​engine for the current move, and use the line-of-sight-move matching coefficient to correct the original AI consistency and generate the behavior attribution consistency; when the line-of-sight-move matching coefficient is invalid, the original AI consistency is directly used as the behavior attribution consistency. S5: Based on the attribution consistency of the behavior and the preset risk classification strategy, determine the cheating risk level of the current move and the game.

[0008] Furthermore, the calculation of the line-of-sight-placement matching coefficient in step S3 includes the following sub-steps: S31: Divide the preset time window into at least two timing stages, the at least two timing stages including a scanning stage and a focusing stage; S32: During the scanning phase, the scanning path complexity is calculated based on the transition characteristics of the gaze area in the gaze area sequence between each preset area of ​​the chessboard. The scanning path complexity is used to measure the degree of global exploration of the candidate's gaze on the chessboard. S33: During the focusing phase, based on the frequency change characteristics of the area where the move coordinates are located in the sequence of the gaze chessboard areas, it is determined whether there is a focusing convergence process. The focusing convergence process is the process in which the examinee's gaze gradually focuses on the area where the move coordinates are located as the move time approaches. S34: Generate the line-of-sight matching coefficient based on the complexity of the scanning path and the presence or absence of the focusing convergence process; wherein, when the complexity of the scanning path does not meet the preset exploration requirements, the line-of-sight matching coefficient is reduced; when the focusing convergence process exists, the line-of-sight matching coefficient is increased.

[0009] Furthermore, before generating the line-of-sight-piece matching coefficient in step S34, the following sub-steps are also included: S3401: During the focusing phase, identify the transition pattern of the gaze entering the area where the piece is placed, the transition pattern including at least a skip-direct approach mode and a progressive convergence mode; wherein, the skip-direct approach mode is that in the frame before entering the area where the piece is placed, the gaze area is not adjacent to the area where the piece is placed; the progressive convergence mode is that in the several frames before entering the area where the piece is placed, the gaze area gradually transitions from an area far from the piece is placed to the area where the piece is placed through adjacent areas. S3402: When the identified transfer mode is the skip-direct mode, a preset transfer mode penalty factor is applied to the line-of-sight matching coefficient to reduce the line-of-sight matching coefficient; when the identified transfer mode is the progressive convergence mode, the line-of-sight matching coefficient is kept unchanged.

[0010] Furthermore, the step S31 of dividing the preset time window into at least two time phases includes the following sub-steps: S311: Calculate the gaze distribution entropy value of the gaze chessboard region sequence at each moment within the preset time window. The gaze distribution entropy value is the information entropy of the probability distribution of each preset region being gazed upon. The higher the gaze distribution entropy value, the more uniform the distribution of the examinee's gaze among the preset regions of the chessboard. S312: Detect the turning point in time when the gaze distribution entropy value continuously decreases from above a preset entropy threshold to below the preset entropy threshold; S313: The detected turning point is used as the first boundary between the scanning phase and the focusing phase, instead of the preset fixed boundary.

[0011] Furthermore, after generating the line-of-sight matching coefficient based on the complexity of the scanning path and the presence or absence of the focusing convergence process in step S34, the following sub-steps are also included: S351: When the complexity of the scanning path is lower than a preset complexity threshold and the focusing convergence process is not detected, the current move is marked as a visually disjointed move. S352: When the number of consecutive visually disjointed moves exceeds a preset disjoint sequence length threshold, a high-risk cheating determination is directly generated, and the high-risk cheating determination is output as the result of the cheating risk level, no longer relying on the line-of-sight matching coefficient and the behavior attribution consistency.

[0012] Furthermore, step S4, which involves using the line-of-sight matching coefficient to correct the original AI fit and generate a behavior attribution fit, includes the following sub-steps: S41: Based on the line-of-sight matching coefficient, the original AI matching degree is split into the part that can be attributed to the candidate's autonomous visual attention and the part that cannot be attributed to the candidate's autonomous visual attention, and autonomous attribution component and non-attribution component are obtained respectively. S42: The unattributed component is used as the behavioral attribution consistency for subsequent cheating risk determination; wherein, the autonomous attribution component is removed from the calculation of the behavioral attribution consistency, so that the high consistency part explained by the visual behavior does not generate cheating suspicion.

[0013] Furthermore, after step S41, based on the gaze-to-placement matching coefficient, decomposes the original AI matching degree into a part attributable to the examinee's autonomous visual attention and a part not attributable to the examinee's autonomous visual attention, and obtains the autonomous attribution component and the non-attribution component respectively, it also includes the following sub-steps: S411: For consecutive multiple moves, construct a joint time series consisting of the line-of-sight-move matching coefficient sequence and the original AI consistency sequence; S412: Slide the detection window on the joint time series. When the increase of the mean value of the original AI matching degree in the detection window compared with the previous window exceeds the preset increase threshold, and the change of the mean value of the line-to-spot matching coefficient in the same detection window is lower than the preset change threshold, it is determined that there is a line-to-spot matching degree deviation anomaly. S413: When it is determined that there is an abnormality in the line-of-sight consistency, a preset deviation penalty factor is applied to the behavior attribution consistency of each piece in the detection window to increase the cheating risk level of the corresponding piece.

[0014] Further, in step S42, the unattributed component is used as the behavioral attribution consistency for subsequent cheating risk determination; wherein, after the autonomous attribution component is removed from the calculation of the behavioral attribution consistency, so that the high consistency portion explained by the visual behavior does not generate cheating suspicion, the following sub-steps are also included: S431: Map the unattributed components of each move in the entire game to the chessboard coordinate system according to their move coordinates, and construct an unattributed spatial distribution map. S432: On the unattributed spatial distribution map, detect whether there are regions on the chessboard where pieces with unattributed components exceeding a preset high value threshold are continuously clustered; the continuously clustered region is a connected region on the chessboard where pieces with high unattributed components exist in adjacent preset regions. S433: When the continuously clustered area is detected and the number of pieces placed in the area exceeds the preset clustering number threshold, a high-risk cheating determination is directly triggered, and the high-risk cheating determination is output as the result of the cheating risk level.

[0015] Further, in step S42, the unattributed component is used as the behavioral attribution consistency for subsequent cheating risk determination; wherein, after the autonomous attribution component is removed from the calculation of the behavioral attribution consistency, so that the high consistency portion explained by the visual behavior does not generate cheating suspicion, the following sub-steps are also included: S441: Obtain the candidate’s personal vision-chess strength baseline constructed by the autonomous attribution components in historical games, wherein the personal vision-chess strength baseline reflects the level of autonomous attribution components corresponding to different board complexities under normal thinking conditions. S442: Compare the autonomous attribution component of the current move with the baseline value corresponding to the current position complexity in the personal vision-chess strength baseline, and calculate the deviation. S443: When the deviation exceeds the preset deviation threshold, the cheating risk level of the current move is increased, even if the unattributed component of the current move has not reached the risk trigger threshold.

[0016] Further, in step S42, the unattributed component is used as the behavioral attribution consistency for subsequent cheating risk determination; wherein, after the autonomous attribution component is removed from the calculation of the behavioral attribution consistency, so that the high consistency portion explained by the visual behavior does not generate cheating suspicion, the following sub-steps are also included: S451: For a continuous sequence of moves, detect whether there is a periodic alternation pattern in which line-of-sight deviation and high-match moves alternate; the line-of-sight deviation is the state in which the line-of-sight-move matching coefficient is lower than a preset deviation threshold, and the high-match moves are moves in which the unattributed component exceeds a preset high-match threshold; the periodic alternation pattern is a pattern in which the line-of-sight deviation and the high-match moves alternate within a period of time and the number of alternations exceeds a preset number of alternations threshold. S452: When the periodic alternation pattern is detected, a high-risk cheating determination is directly triggered, and the high-risk cheating determination is output as the result of the cheating risk level, instead of relying on the attribution consistency of the behavior of a single move.

[0017] Accordingly, a second aspect of the present invention provides an electronic device, including: at least one processor; and a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the visual detection anti-cheating method of the chess interactive device described above.

[0018] Accordingly, a third aspect of the present invention provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the visual detection anti-cheating method of the above-described chess-playing interactive device.

[0019] The above-described technical solutions of the embodiments of the present invention have the following beneficial technical effects: 1. By decomposing the candidate's gaze behavior before placing a piece into a scanning phase and a focusing phase, and extracting the scanning path complexity and focusing convergence features, this invention can restore the complete visual thinking chain of "global observation - local focusing - confirmation of placement". This improves the system's evaluation of the candidate's gaze behavior from the static duration statistics of the existing technology to dynamic temporal structure analysis, thereby effectively distinguishing between the two behavior modes of "placing a piece after sufficient observation and consideration" and "placing a piece directly without observing the chessboard", which significantly improves the accuracy of cheating behavior identification and the interpretability of the judgment results. 2. By splitting the original AI consistency score into an autonomous attribution component and an unattribution component, and removing the autonomous attribution component from the cheating risk calculation, this invention enables the consistency score index to carry behavioral attribution information. The high consistency score generated by high-level players under normal thinking conditions is automatically classified into the autonomous attribution component and is exempt from triggering alarms. Only the unattribution component, which lacks visual support, is used as an effective risk input. This fundamentally solves the inherent defect in existing AI consistency score analysis schemes where strong players are easily falsely accused, and significantly reduces the false alarm rate. 3. By further introducing joint time-series deviation detection, chessboard space clustering analysis, individual line-of-sight-chess strength baseline deviation detection, and line-of-sight deviation-high-match periodic alternation pattern detection on the basis of semantic segmentation of consistency, this invention constructs a complete system for three-dimensional identification of cheating behavior from multiple dimensions of time series, space, individual baseline, and behavioral rhythm. This enables decentralized cheating behaviors with high concealment and weak single-instance evidence to be locked by multi-dimensional cross-validation, significantly improving the identification coverage and judgment robustness of complex cheating methods in online examination scenarios. Attached Figure Description

[0020] Figure 1 This is a flowchart of the visual detection anti-cheating method for a chess-playing interactive device provided in this embodiment of the invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0022] This invention provides a visual detection anti-cheating method for a chess-playing interactive device, applicable to scenarios requiring fairness in remote Go matches, such as online Go ranking exams, online tournaments, and remote teaching assessments by training institutions. In these scenarios, examinees or players are typically in private, non-uniformly supervised environments, making it difficult to effectively monitor their visual attention behavior and surrounding environment using only a self-installed single camera device. This method operates within a chess-playing interactive device, which includes at least a first image acquisition module pointing towards the examinee's face, a second image acquisition module pointing towards the chessboard area, a local processing unit, and a network module communicating with an AI engine. The first image acquisition module continuously acquires images of the examinee's face and upper body, providing input data for subsequent calculation of the gaze area; the second image acquisition module continuously acquires chessboard images from a top-down angle, used to identify changes in piece positions and detect move events; the local processing unit performs real-time processing tasks such as gaze area calculation, gaze area sequence analysis, gaze-move matching coefficient calculation, and behavior attribution consistency generation; the network module sends real-time game records to the AI ​​engine and receives the returned raw AI consistency data. In practical deployment, this chess-playing interactive device typically includes a standard Go board and a matching move-making mechanism. The Go board is fixed to the top of the device. A first image acquisition module is mounted on the top of the device's pillar to capture the viewing angle facing the examinee, and a second image acquisition module is mounted on a bracket directly above the Go board to ensure complete coverage of the Go board area. The Go board plane is pre-divided into multiple preset Go board areas. The division can be done using a uniform grid, for example, dividing a 19x19 Go board into several rectangular areas using a 4x4 or 6x6 grid. Each preset Go board area covers several intersections on the Go board, and this division serves as the reference basis for mapping the gaze area to the Go board coordinate system. The method of this invention will be described in detail below with specific steps.

[0023] Please refer to Figure 1 The first aspect of this invention provides a visual detection anti-cheating method for a chess-playing interactive device, comprising the following steps: S1: The candidate's facial image is acquired by a first image acquisition module pointing towards the candidate's face. The candidate's gaze area is calculated in real time and mapped to a pre-calibrated chessboard coordinate system to obtain a gaze chessboard region sequence. The gaze chessboard region is the area where the mapped point of the gaze area falls on the chessboard plane after dividing the chessboard plane into multiple preset regions.

[0024] The first image acquisition module is typically mounted on top of the pillar or above the display of the chess-playing interactive device, covering the upper body and face of the examinee from a top-down or eye-level angle. After the game begins, the first image acquisition module captures facial images of the examinee in consecutive frames. The local processing unit runs a face detection model on each frame to locate the facial region and extracts facial feature points within the detected region, including key points such as the eyes, nose tip, and corners of the mouth. Based on the extracted facial feature points, a head pose calculation algorithm is used to calculate the three-dimensional rotation angles of the head, including pitch, yaw, and roll angles. In a simplified implementation, it can be assumed that the examinee's eyes are centered and the direction of their gaze is consistent with the direction of the facial normal, thus constructing the gaze direction vector from the head rotation angles.

[0025] To translate the gaze direction into a meaningful physical location on the chessboard, the calibration between the first image acquisition module and the chessboard coordinate system must be completed beforehand. The calibration process can be completed at the factory or via a guided program during the student's first use. The specific calibration method is as follows: A chessboard world coordinate system is established with the top-left corner of the chessboard as the origin, the row direction as the horizontal axis, and the column direction as the vertical axis. The extrinsic parameter matrix of the first image acquisition module relative to the chessboard world coordinate system is calculated using reference points with known world coordinates on the chessboard. During actual gameplay, a spatial ray equation is constructed based on the gaze direction vector and the camera's optical center position. The intersection of this ray with the chessboard plane (i.e., the plane with zero height in the chessboard world coordinate system) is solved to obtain the coordinates of the student's gaze point on the chessboard at the current moment. Because head pose estimation inherently involves some noise, this step does not require the gaze point to be precise down to the level of a single chessboard intersection. Instead, the chessboard plane is pre-divided into several grid regions, such as a four-row, four-column or six-row, six-column layout. The gaze point coordinates calculated for each frame are assigned to the corresponding grid region, forming a gaze chessboard region sequence over time. This sequence records which pre-defined region of the chessboard the examinee's gaze falls on at each moment during the entire game, providing structured input data for the temporal analysis in subsequent steps.

[0026] S2: Detect the move event by using the second image acquisition module pointing to the chessboard area, obtain the coordinates of the current move in the chessboard coordinate system, and record the move time.

[0027] The second image acquisition module is typically mounted above the chessboard at a top-down angle to continuously acquire images of the chessboard area. After perspective correction and chessboard grid localization, changes in the chessboard state are detected using frame difference or a deep learning-based piece segmentation model. When the examinee places a new piece on the chessboard, the second image acquisition module identifies the new piece and extracts its pixel position in the chessboard image. Using the established mapping between the calibrated chessboard image coordinate system and the chessboard world coordinate system, the pixel position is converted into the placement coordinates in the chessboard coordinate system. Simultaneously, the precise moment of the placement event is recorded as the placement time. This placement time serves as the reference point for subsequent extraction of the gaze behavior time window. During the examinee's placement of a piece, the robotic arm or the examinee's arm may briefly occlude the chessboard area. The second image acquisition module re-detects the chessboard state after the occlusion is removed to ensure the accuracy of the placement detection.

[0028] S3: For the current move, extract the gaze board region sequence within a preset time window before the move. Calculate the gaze-move matching coefficient based on the proportion of valid gaze frames matching the move's coordinates within the gaze board region sequence to the total number of valid gaze frames within the preset time window. When the proportion of valid gaze frames within the preset time window is lower than a preset threshold, the gaze-move matching coefficient is marked as invalid.

[0029] Once a move is detected, the time of that move is used as the time reference point. A preset time window is then traced back, and the sequence of gaze areas on the chessboard generated by S1 within that time window is extracted. The length of the time window can be set according to the typical human thinking rhythm in Go, usually a few seconds, to cover the key period when the player observes and makes decisions before making a move.

[0030] Within the extracted time window, the total number of valid gaze frames in the gaze board region sequence is counted. A valid gaze frame is defined as a frame where the gaze direction has a valid intersection with the chessboard plane and the gaze falls within the chessboard area, excluding frames where the gaze deviates from the chessboard or where detection fails. Simultaneously, the number of valid gaze frames that match the gaze region with the preset region where the current move coordinates are located is counted. Dividing the number of matching frames by the total number of valid gaze frames yields the matching ratio between the gaze region and the move region within the time window, i.e., the gaze-move matching coefficient. This coefficient reflects the degree of visual attention the examinee paid to the move region before making their move: a higher value indicates that the examinee continuously gazed at the final move region before making their move, and the move was supported by sufficient visual observation; a lower value indicates that the examinee paid less attention to the region before making their move, and the move lacked a causal basis for visual attention.

[0031] In real-world applications, candidates may be unable to effectively detect gaze direction for a certain period due to factors such as looking down, covering their face with their arms, or deliberately avoiding the camera. To address this, a validity assessment mechanism is implemented in this step: when the proportion of valid gaze frames within a time window to the total number of frames expected in that window is lower than a preset threshold, it indicates that the overall reliability of the gaze data within that window is insufficient. In this case, the calculated gaze-placement matching coefficient is marked as invalid and not used for subsequent correction. This mechanism ensures the reliability of the gaze-placement matching coefficient and avoids misleading judgments due to low data quality.

[0032] S4: Obtain the original AI consistency score calculated by the AI ​​engine for the current move, and use the line-of-sight-move matching coefficient to correct the original AI consistency score, generating the behavior attribution consistency score. When the line-of-sight-move matching coefficient is invalid, the original AI consistency score is directly used as the behavior attribution consistency score.

[0033] The AI ​​engine, deployed in the cloud or on a local server, is a self-developed Go-playing AI engine. When S2 detects a move, the local processing unit sends the corresponding real-time game record to the AI ​​engine via the network module. The AI ​​engine calculates the AI's recommended move and its probability distribution based on the current board position and returns a score indicating the degree of consistency between the move and the AI's preferred move or one of its top choices—the original AI consistency score. The original AI consistency score can be calculated as follows: a high value is assigned if the move perfectly matches the AI's preferred move; a medium value is assigned if the move falls within the AI's top five choices, decreasing in ranking; and a low or zero value is assigned if the move is outside the recommended range.

[0034] After obtaining the initial AI consistency score, the line-of-sight matching coefficient calculated by S3 is used to correct the initial AI consistency score, generating a behavioral attribution consistency score. The basic logic of the correction is: the line-of-sight matching coefficient is used as a causal attribution factor to affect the initial AI consistency score. When the line-of-sight matching coefficient is high, it indicates that the area where the move was made received continuous visual attention from the examinee before the move was made. This move can be attributed to the examinee's independent observation and thinking. Even if the initial AI consistency score is high, its behavioral attribution consistency score should be lowered accordingly to reduce the suspicion of cheating. When the line-of-sight matching coefficient is low, it indicates that the examinee did not pay sufficient attention to the area before the move was made, and the move selection cannot be explained by independent visual attention. The suspicion of cheating should be retained or amplified, and the behavioral attribution consistency score should be kept high or further increased.

[0035] In one specific correction method, the original AI consistency score can be multiplied by a correction weight determined by the line-of-sight matching coefficient, which decreases as the matching coefficient increases. When S3 has already marked the line-of-sight matching coefficient as invalid, this step abandons the correction operation and directly uses the original AI consistency score as the behavior attribution consistency score. This allows the method to automatically degenerate into a judgment method that relies solely on the AI ​​consistency score when the line-of-sight data is unreliable, ensuring the normal operation of the system under any conditions.

[0036] S5: Based on the consistency of behavioral attribution and the preset risk classification strategy, determine the cheating risk level of the current move and the game.

[0037] Based on the behavioral attribution consistency of each move generated by S4, combined with a preset risk grading strategy, the risk level of the current move is determined. Furthermore, it can integrate the behavioral attribution consistency of multiple moves throughout the entire game to provide an overall game-level cheating risk assessment. The risk grading strategy defines the mapping relationship between behavioral attribution consistency values ​​and risk levels. For example, it sets low-risk, medium-risk, and high-risk ranges. When the behavioral attribution consistency of a move falls into the high-risk range, an immediate alarm is triggered. When the cumulative number of medium-risk events reaches a certain number, it is escalated to high-risk processing. Low-risk events are only recorded and do not interfere with the exam. Different risk levels correspond to differentiated response actions: low-risk events do not interrupt the exam process; medium-risk events remind candidates to pay attention to exam discipline through voice prompts or screen pop-ups; and high-risk events automatically notify the referee team and upload relevant invigilation video clips.

[0038] Through the above steps, this embodiment incorporates the real-time visual gaze behavior of examinees during the game into the AI ​​consistency evaluation system. This transforms the AI ​​consistency index from an isolated numerical value into a causal relationship established with the examinee's gaze behavior. Gazing data, as behavioral evidence, directly participates in the correction of chess strength analysis indicators, enabling the final consistency index used to determine cheating risk to distinguish between "normally high consistency generated through independent observation and thinking" and "abnormally high consistency lacking visual support." This fundamentally changes the existing model where visual behavior monitoring and AI consistency analysis operate independently, achieving deep integration of behavioral evidence and chess strength analysis at the data level. This significantly improves the accuracy and fairness of cheating determination in online examination scenarios.

[0039] Specifically, step S3 calculates the line-of-sight-piece matching coefficient, including the following sub-steps: S31: Divide the preset time window into at least two timing phases, including a scanning phase and a focusing phase.

[0040] The preset time window is a fixed time interval preceding the move, covering the candidate's complete cognitive process from observing the board to making a move. The scanning phase corresponds to the earlier part of the time window, during which the candidate typically observes the entire board broadly, shifting their gaze between different areas to gather information. The focusing phase corresponds to the period closer to the move, during which the candidate's attention gradually converges to the final move area, performing a confirmatory gaze before placing the move. The first dividing point between the two phases can be determined according to a preset fixed ratio, such as designating the first half of the time window as the scanning phase and the second half as the focusing phase; alternatively, it can be dynamically divided based on the statistical characteristics of the gaze area distribution to accommodate differences in thinking pace among different candidates and in different board situations.

[0041] S32: During the scanning phase, the saccade path complexity is calculated based on the transition characteristics of the gaze region in the gaze region sequence between each preset region of the chessboard. The saccade path complexity is used to measure the degree of global exploration of the candidate's gaze on the chessboard.

[0042] Sagging path complexity is a comprehensive measure of the range and frequency of a candidate's gaze shifts between various pre-defined areas of the chessboard. It is calculated based on the number of times the gaze area changes during the scanning phase and the number of different pre-defined chessboard areas covered. A change in gaze area refers to the shift of the gaze area from one pre-defined area to another between adjacent frames; the number of changes reflects the frequency of the candidate's swagging; the number of different pre-defined areas covered reflects the breadth of the candidate's swagging range. The swagging path complexity is obtained by comprehensively calculating the number of swagging shifts and the number of covered areas. A higher value indicates a more thorough exploration of the entire chessboard during the scanning phase, providing a stronger basis for the candidate's subsequent move decisions through independent observation.

[0043] S33: During the focusing phase, based on the frequency change characteristics of the areas where the coordinates of the pieces are located in the sequence of areas on the chessboard, determine whether there is a focusing convergence process. The focusing convergence process is the process by which the examinee's gaze gradually focuses on the area where the coordinates of the pieces are located as the time of the piece placement approaches.

[0044] The convergence process refers to the process by which a player's gaze gradually focuses on the area where the move is located as the move approaches. Specifically, the sequence of gazed areas on the chessboard during the focusing phase is analyzed chronologically, calculating the frequency of occurrence of the area where the move is located in each time segment, and observing whether this frequency shows a gradual increase as the move approaches. When the frequency shows a monotonically increasing trend from low to high, eventually stabilizing at the area where the move is located, it can be determined that a convergence process exists. The existence of this convergence process indicates that the player has undergone a confirmatory gaze phase of "locking onto the target area" before making the move, and this move conforms to the natural behavioral pattern of "observation-focusing-move" in human Go play.

[0045] S34: Generate a line-of-sight matching coefficient based on the complexity of the scanning path and the presence or absence of a focusing convergence process. Specifically, when the scanning path complexity does not meet the preset exploration requirements, the line-of-sight matching coefficient is decreased. When a focusing convergence process exists, the line-of-sight matching coefficient is increased.

[0046] The line-of-sight matching coefficient is a comprehensive indicator that quantifies the degree to which a candidate's visual attention supports their move before it is made. When the scanning path complexity is low and fails to meet the preset exploration requirements, it indicates that the candidate has not fully observed the entire board before making a move, and the foundation for independent thinking in making that move is weak. In this case, the value of the line-of-sight matching coefficient should be reduced to increase the weight of the move as suspected of cheating. When a focusing convergence process is detected during the focusing phase, it indicates that the candidate's line of sight naturally converges to the target area before making a move, forming a complete behavioral chain of "observation-focusing-movement". In this case, the value of the line-of-sight matching coefficient should be increased to reduce the weight of the move as suspected of cheating. The preset exploration requirements for scanning path complexity and the judgment threshold for the focusing convergence process can be set and adjusted according to the actual application scenario to adapt to the differences in behavioral characteristics of candidates at different levels and age groups.

[0047] The above technical solution decomposes the visual attention process before placing a piece into a scanning stage and a focusing stage, and performs feature extraction and comprehensive judgment on each stage. This makes the gaze-placing matching coefficient reflect not only the static result of "whether the examinee gazed at the placement area" but also the dynamic process of "how the examinee gazed at the placement area." This provides matching parameters with behavioral explanation capabilities for the subsequent causal correction of AI matching, significantly improving the precision and accuracy of visual behavior analysis in the anti-cheating system.

[0048] Furthermore, before generating the line-of-sight-piece matching coefficients in step S34, the following sub-steps are also included: S3401: During the focusing phase, identify the transition pattern of the gaze entering the area where the piece's coordinates are located. The transition pattern includes at least a skip-direct approach pattern and a progressive convergence pattern. The skip-direct approach pattern occurs when, in the frame preceding the entry into the area where the piece's coordinates are located, the gaze area is not adjacent to the area where the piece's coordinates are located. The progressive convergence pattern occurs when, in the preceding frames, the gaze area gradually transitions from an area far from the piece's coordinates to the area where the piece's coordinates are located, passing through adjacent areas.

[0049] After the scanning phase, the candidate's gaze shifts from a certain area of ​​the chessboard to the final placement area. This shift is not uniform but exhibits different behavioral characteristics. This invention distinguishes the shift patterns into at least two types: a jump-direct approach mode and a progressive convergence mode. The jump-direct approach mode refers to a situation where, in the frame preceding the placement coordinates area, the pre-defined chessboard area where the gaze is focused is not adjacent to the placement coordinates area. In other words, the gaze jumps directly from an area on the chessboard that is not directly adjacent to the placement area to the placement area, without passing through adjacent areas. This shift method is relatively rare in actual autonomous thinking processes because players usually conduct gradual visual confirmation of the placement area and its surroundings before making a move decision, rather than jumping directly from a distant area. The progressive convergence mode refers to a situation where, in the several frames preceding the placement coordinates area, the gaze starts from an area far from the placement coordinates, sequentially passes through one or more pre-defined chessboard areas adjacent to the placement area, gradually transitions, and finally enters the placement coordinates area. This gradually converging transfer path conforms to the natural behavioral patterns of human visual search and decision confirmation, indicating that the examinee underwent an observation and confirmation process from far to near before placing the piece.

[0050] S3402: When the identified transfer mode is a jump-direct mode, a preset transfer mode penalty factor is applied to the line-of-sight matching coefficient to reduce the line-of-sight matching coefficient. When the identified transfer mode is a progressive convergence mode, the line-of-sight matching coefficient is kept unchanged.

[0051] When the identified transfer pattern is a jump-direct approach, it indicates that the candidate's gaze suddenly locks onto the placement area without a gradual confirmation process. This behavior pattern highly matches the behavior of directly locating the target placement point with the help of external cues. Therefore, a preset transfer pattern penalty factor is applied to the gaze-placement matching coefficient of this move, reducing the value of the matching coefficient and increasing the suspicion of cheating. The transfer pattern penalty factor is a coefficient less than 1, and its specific value can be set according to the correlation strength between the jump-direct approach pattern and cheating behavior in the actual scenario. When the identified transfer pattern is a gradual convergence pattern, it indicates that the candidate's gaze transfer process conforms to the behavioral pattern of autonomous observation and thinking. In this case, the gaze-placement matching coefficient remains unchanged, and no additional penalty or reward adjustment is applied.

[0052] Furthermore, step S31 divides the preset time window into at least two timing stages, including the following sub-steps: S311: Calculate the gaze distribution entropy value of the gaze chessboard region sequence at each moment within a preset time window. The gaze distribution entropy value is the information entropy of the probability distribution of gaze in each preset region. The higher the gaze distribution entropy value, the more uniform the distribution of the examinee's gaze among the preset regions of the chessboard.

[0053] The gaze distribution entropy is an application of the concept of entropy in information theory to the field of gaze behavior analysis. It is used to quantify the uniformity of a candidate's gaze distribution across various pre-defined areas of a chessboard. Within a pre-defined time window, at a specific moment or within a short time segment, the frequency of the candidate's gaze falling into each pre-defined area of ​​the chessboard is statistically analyzed to construct a probability distribution of the gaze area across each pre-defined area, and then the information entropy of this probability distribution is calculated. When the gaze area is concentrated in a few pre-defined areas of the chessboard or even a single area, the probability distribution is concentrated, and the gaze distribution entropy is low; when the gaze area is relatively evenly distributed across the pre-defined areas of the chessboard, the probability distribution is dispersed, and the gaze distribution entropy is high. Therefore, the lower the gaze distribution entropy, the more concentrated the candidate's gaze is in a local area, exhibiting a focused state; the higher the gaze distribution entropy, the more evenly the candidate's gaze is distributed across the entire chessboard, exhibiting a scanning state. In practical calculations, the gaze chessboard area sequence within a pre-defined time window can be segmented and sampled at pre-defined time intervals. The distribution probability of the gaze area is statistically analyzed and the entropy value is calculated for each sample segment, resulting in a sequence of gaze distribution entropy values ​​arranged chronologically.

[0054] S312: Detect the turning point when the gaze distribution entropy value continuously decreases from above the preset entropy threshold to below the preset entropy threshold along the time sequence.

[0055] In the normal human Go thinking process, players typically first scan the entire board broadly, at which point the gaze distribution entropy value remains at a high level. As thinking deepens and the intention to place a move becomes clearer, the gaze gradually focuses on at least a few candidate areas, and the gaze distribution entropy value decreases accordingly. Near the moment of placing a move, the gaze finally converges on the move area, and the gaze distribution entropy value drops to its lowest level. This process of "entropy value continuously decreasing from high to low" reflects the player's natural transition from the global scanning stage to the local focusing stage. S312 traverses the gaze distribution entropy value sequence chronologically. When it detects that the entropy value continuously decreases from a level above a preset entropy threshold and then falls below the preset entropy threshold at a certain moment, that moment is marked as a turning point. The judgment of continuous decrease can be achieved by requiring the entropy value of several consecutive sampling points to decrease monotonically, thus avoiding misjudgments caused by entropy fluctuations in individual noisy frames.

[0056] S313: The detected turning point is used as the first boundary between the scanning phase and the focusing phase, instead of the preset fixed boundary.

[0057] Using the dynamic segmentation method based on gaze distribution entropy, the scanning phase is defined as the time period from the start of the preset time window to the first boundary time, and the focusing phase is defined as the time period from the first boundary time to the placement of the piece. This adaptive segmentation method allows the phase division results to automatically adapt to individual differences among test takers. Test takers with slower thinking paces naturally have longer scanning phases, while those with faster thinking paces have shorter scanning phases. It also adapts to differences in the complexity of different situations; test takers require longer global scanning time in complex situations, while they can enter the focusing phase more quickly in simple situations. Compared with the method using fixed boundary times, the adaptive segmentation based on gaze distribution entropy significantly improves the accuracy and rationality of phase division, thereby enhancing the complexity of the saccade path and the effectiveness of focusing convergence feature extraction, making the gaze-placing matching coefficient more accurately reflect the test taker's true visual attention behavior pattern.

[0058] Furthermore, after generating the line-of-sight matching coefficients in step S34 based on the complexity of the scanning path and the presence or absence of the focusing convergence process, the following sub-steps are also included: S351: When the scanning path complexity is lower than the preset complexity threshold and no focusing convergence process is detected, the current move is marked as a visually disjointed move.

[0059] Visual disconnect in a move refers to a situation where, within a pre-defined time window before the move is made, the candidate's visual attention fails to demonstrate either a thorough scan of the entire board or a gradual convergence towards the move's location. Specifically, when the scanning path complexity is below a pre-defined threshold, it indicates that the candidate has not adequately explored all areas of the board during the scanning phase, exhibiting a narrow visual range, low scanning rate, and a lack of observation of the overall situation. Simultaneously, when no convergence is detected during the focusing phase, it suggests that the candidate's gaze does not gradually converge towards the move's location as the move approaches, indicating a lack of pre-positioning visual attention. When both visual deficiencies occur simultaneously, there is a complete disconnect between the candidate's visual attention and the move decision. This move lacks both a global observation basis and a local focusing confirmation process, and its visual characteristics are highly consistent with cheating patterns such as "making a move without looking at the board" or "making a move directly with external prompts." Therefore, this move is marked as a visual disconnect. The marking of visually disjointed moves is independent of the degree of AI matching. Regardless of whether the move's position matches the AI's recommended choice, it is marked as long as the visual attention behavior meets the above two missing conditions.

[0060] S352: When the number of consecutive visually disjointed moves exceeds the preset disjoint sequence length threshold, a high-risk cheating judgment is directly generated, and the high-risk cheating judgment is output as the result of the cheating risk level, instead of relying on the line-to-movement matching coefficient and behavior attribution consistency.

[0061] A single instance of visual disconnection may be caused by chance, such as a candidate making a quick move without a complete observation and thought process when the position is extremely simple and the only possible move is available. However, when visual disconnections occur consecutively, the explanatory power of chance decreases significantly, and the lack of visual attention support for multiple consecutive moves constitutes a strong signal of cheating. A preset threshold for the length of the disconnection sequence is used to define the boundary between the tolerable number of occasional visual disconnections and the systematic visual disconnection behavior that should trigger an alarm. When the number of visual disconnections in a consecutive move sequence exceeds this threshold, the system directly generates a high-risk cheating judgment and outputs this high-risk cheating judgment as the result of the cheating risk level. This judgment path is completely independent of the line-of-sight matching coefficient and the subsequent behavioral attribution consistency calculation. Even if the AI ​​consistency value of the visual disconnection is low or the behavioral attribution consistency does not reach the risk trigger threshold, the fact of consecutive visual disconnections itself constitutes sufficient evidence for cheating identification.

[0062] The aforementioned visual disjointed move sequence determination mechanism elevates visual attention behavior from an auxiliary correction role for AI consistency to an independent source of cheating evidence: when the visual behavior itself exhibits a clear abnormal pattern, there is no need to rely on the accumulation and analysis of AI consistency values; a determination can be made directly based on the facts of the visual behavior. This mechanism significantly enhances the anti-cheating system's ability to capture the extreme cheating pattern of "completely independent of autonomous observation," while providing highly interpretable behavioral evidence to support cheating determination.

[0063] Specifically, step S4 uses the line-of-sight-piece matching coefficient to correct the original AI fit and generate the behavior attribution fit, including the following sub-steps: S41: Based on the line-of-sight matching coefficient, the original AI matching degree is split into the part that can be attributed to the candidate's autonomous visual attention and the part that cannot be attributed to the candidate's autonomous visual attention, and autonomous attribution component and non-attribution component are obtained respectively.

[0064] The original AI consistency score, calculated by the AI ​​engine based on the current board position, represents the degree of agreement between the current move and the AI's recommended move in that position; it is a uniform numerical indicator. However, the same consistency score may correspond to completely different behavioral contexts: a high consistency score may stem from a high-quality move independently derived by the player after sufficient independent observation and in-depth thinking, or it may stem from move suggestions obtained by the player using external AI tools without observing the board. In existing technologies, the AI ​​consistency score cannot distinguish between these two situations, making it easy for the normal performance of high-level players to be misjudged as cheating. This invention introduces a line-of-sight matching coefficient as an attribution basis, decomposing the original AI consistency score into two components with clear behavioral implications. The autonomous attribution component represents the portion of the original AI consistency score that can be explained by the player's visual attention behavior. Its value is jointly determined by the original AI consistency score and the line-of-sight matching coefficient. The higher the line-of-sight matching coefficient, the more sufficient the player's visual attention to the move area is before making the move, and the more likely the consistency score of that move stems from independent thinking; therefore, a higher proportion is attributed to the autonomous attribution component. The unattributed component represents the remaining portion of the original AI consistency that cannot be explained by the test taker's visual attention behavior; that is, the consistency value lacking visual behavior support. For example, when the original AI consistency for a move is high but the visual-placement matching coefficient is low, it indicates that the test taker made a move highly consistent with the AI ​​recommendation without fully observing the placement area. In this case, the autonomous attribution component is small and the unattributed component is large. Conversely, when the visual-placement matching coefficient is high, even if the original AI consistency is high, most of the consistency will be attributed to the autonomous attribution component, and the unattributed component will be small.

[0065] S42: The unattributed component is used as the behavioral attribution consistency score for subsequent cheating risk assessment. Specifically, the autonomous attribution component is removed from the behavioral attribution consistency score calculation, ensuring that the high consistency score explained by visual behavior does not raise suspicion of cheating.

[0066] Behavioral attribution consistency is a consistency index adjusted for causal factors related to visual behavior. Its core feature is the complete removal of autonomous attribution components from risk calculation, retaining only the unattributed components that cannot be explained by visual behavior as valid risk input. This approach reflects the core judgment principle that distinguishes this invention from existing technologies: high consistency explained by visual behavior does not raise suspicion of cheating. In actual games, high-level players often make moves that are highly consistent with AI-recommended moves after long periods of observation and in-depth thinking. However, this high consistency is supported by sufficient visual attention behaviors such as scanning exploration and focusing convergence. Therefore, most of the consistency is attributed to autonomous attribution components and removed, keeping behavioral attribution consistency at a low level and not triggering cheating alerts. Conversely, if a candidate makes a high-consistency move when their gaze is off the board, the scanning path complexity is extremely low, or there is a lack of focusing convergence, this consistency cannot be explained by visual behavior, and most of it enters the unattributed components. Behavioral attribution consistency increases significantly, thus triggering a corresponding level of cheating risk judgment. When the gaze-to-pick matching coefficient is marked as invalid, for example, due to the low proportion of valid gaze frames within the preset time window leading to unreliable gaze data, the system cannot perform meaningful attribution decomposition on the original AI consistency. In this case, the original AI consistency is directly used as the behavior attribution consistency, which is handled in a conservative manner to avoid missing potential cheating behaviors due to missing gaze data.

[0067] By splitting the original AI consistency score into autonomous attribution and non-attribution components and removing the former from risk calculation, this invention achieves a fundamental shift in the AI ​​consistency score index from "quantitative comparison" to "semantic attribution." This enables the anti-cheating system to automatically distinguish between "high consistency scores generated through autonomous observation and thinking" and "abnormally high consistency scores lacking visual support." This fundamentally solves the inherent flaw in existing AI consistency score analysis schemes where high-level players are easily misreported, significantly improving the accuracy and fairness of cheating determination.

[0068] Furthermore, in step S41, based on the gaze-to-placement matching coefficient, the original AI matching degree is broken down into a part attributable to the examinee's autonomous visual attention and a part not attributable to the examinee's autonomous visual attention. After obtaining the autonomous attribution component and the non-attribution component, the following sub-steps are also included: S411: For consecutive multiple moves, construct a joint time series consisting of the line-of-sight-move matching coefficient sequence and the original AI consistency sequence.

[0069] The line-of-sight matching coefficient sequence is composed of the line-of-sight matching coefficients for each move, arranged in chronological order, reflecting the changing trend of the player's visual attention behavior as the game progresses. The original AI consistency sequence is composed of the original AI consistency scores for each move, arranged in chronological order, reflecting the changing trend of consistency between the player's move quality and the AI ​​recommendation. Aligning the two sequences in order of move placement creates a joint time series, which includes temporal changes in both visual behavior and skill performance dimensions. In practice, the joint time series is updated in real-time as the game progresses. Whenever a new move is detected and the corresponding line-of-sight matching coefficient and original AI consistency score are calculated, the move's data is appended to the end of the joint time series, providing a continuously updated data foundation for subsequent sliding window detection.

[0070] S412: Slide the detection window on the joint time series. When the increase of the mean of the original AI matching degree in the detection window compared with the previous window exceeds the preset increase threshold, and the change of the mean of the line-to-spot matching coefficient in the same detection window is lower than the preset change threshold, it is determined that there is an abnormal line-to-spot matching degree deviation.

[0071] The sliding detection window covers a time window of several consecutive moves, and its length can be set according to the sensitivity requirements for detecting cheating behavior in the actual scenario. During the detection process, the average value of the original AI matching degree within the current detection window is compared with the average value of the preceding window, which is a sequence of moves of equal length adjacent to the current detection window and preceding it in time. When the average value of the original AI matching degree within the current detection window shows a significant increase compared to the preceding window, exceeding a preset increase threshold, it indicates that the candidate's move quality has suddenly improved in this stage, and the consistency between their moves and the AI-recommended selection points is significantly higher than in the previous stage. At the same time, the change in the average value of the line-of-sight matching coefficient within the current detection window compared to the preceding window is detected. When the change in the average value of the line-of-sight matching coefficient is lower than a preset change threshold, it indicates that the candidate's visual attention behavior has not changed significantly in this stage, and the gaze area distribution, saccade path complexity, and focus convergence characteristics are basically the same as in the previous stage. When both of the above conditions are met simultaneously, the original AI's consistency increases significantly while the line-of-sight and move-matching coefficients remain essentially unchanged, an anomaly in the line-of-sight-consistency is identified. The essence of this anomaly is that the candidate's chess skills suddenly improve while their visual attention behavior does not change synchronously, suggesting that the candidate may have obtained additional move-matching decision information without relying on visual observation. This behavioral pattern highly matches the cheating characteristic of "starting to use off-field AI assistance at a certain moment."

[0072] S413: When an abnormality in line-of-sight consistency is detected, a preset deviation penalty factor is applied to the behavior attribution consistency of each piece within the detection window to increase the cheating risk level of the corresponding piece.

[0073] The deviation penalty factor is a coefficient greater than 1. Its function is to proportionally amplify the attribution consistency of each move within the detection window, thereby increasing the cheating risk level of these moves. The logic behind applying the deviation penalty factor is that the detection results of line-of-sight deviation anomalies provide additional temporal evidence of cheating behavior for the moves within the detection window. These moves may not only have an excessively high unattributed component, but also exhibit an abnormal characteristic of a sudden increase in playing strength without an increase in visual attention throughout the phase. This cross-move trend evidence strengthens the suspicion of cheating for moves within that phase. Therefore, by adjusting the attribution consistency of behavior through the deviation penalty factor, the risk level of the moves in that phase reflects the additional suspicion weight of their temporal anomalies. The specific value of the deviation penalty factor can be set and adjusted according to the strength of the correlation between line-of-sight deviation anomalies and cheating behavior in practical applications.

[0074] The aforementioned line-of-sight-matching anomaly detection mechanism extends anti-cheating analysis from single-point judgment on each move to trend analysis of continuous move sequences, enabling effective identification of the dynamic behavior pattern of "cheating starting midway." This mechanism, in conjunction with line-of-sight-matching coefficient analysis based on single-hand moves and semantic decomposition of matching degree, allows the anti-cheating system to capture both static anomalies of single-hand moves and dynamic anomalies of cross-hand moves, significantly enhancing the coverage and accuracy of cheating behavior identification from a temporal perspective.

[0075] Furthermore, in step S42, the unattributed component is used as the behavioral attribution consistency score for subsequent cheating risk assessment. After removing the autonomous attribution component from the behavioral attribution consistency score calculation, ensuring that the high consistency score explained by visual behavior does not raise suspicion of cheating, the following sub-steps are also included: S431: Map the unattributed components of each move in the entire game to the chessboard coordinate system according to the move coordinates, and construct an unattributed spatial distribution map.

[0076] Throughout the game, each move, after semantic decomposition of the matching degree, generates a corresponding unattributed component value. This value reflects the portion of the AI ​​matching degree that cannot be explained by the examinee's visual attention behavior. The unattributed component of each move is associated with its coordinates on the board and recorded. This association accumulates gradually as the game progresses, forming a spatial distribution of unattributed components covering all areas of the board. The board coordinate system is the aforementioned pre-defined coordinate system, where each intersection on the board plane corresponds to a unique coordinate position. In actual construction, the board coordinate system can be overlaid with a pre-defined grid of areas. The unattributed components of moves occurring within each pre-defined area of ​​the board can be aggregated or averaged to form an unattributed spatial distribution map based on the pre-defined area of ​​the board. This distribution map visually demonstrates the spatial distribution pattern of "AI matching degree that cannot be explained by visual behavior" on the board throughout the entire game.

[0077] S432: On the unattributed spatial distribution map, detect whether there are regions on the chessboard where moves with unattributed components exceeding a preset high threshold are continuously clustered. A continuously clustered region is a connected region on the chessboard where both adjacent preset regions contain moves with high unattributed components.

[0078] High unattributed component moves refer to moves whose unattributed component exceeds a preset high threshold. Such moves indicate a high proportion of AI-generated match that cannot be explained by visual behavior, raising strong suspicion of cheating. In the entire game, isolated high unattributed component moves may be due to chance, such as a move being a standard response in the current situation, requiring minimal observation from the player. However, if multiple high unattributed component moves exhibit a continuous clustering on the board, the explanatory power of chance significantly decreases. A continuously clustered region refers to a connected region on the board where high unattributed component moves exist in adjacent preset regions. This region consists of multiple adjacent preset regions, each containing at least one high unattributed component move. This spatial clustering pattern closely matches the typical characteristics of human cheating behavior: cheaters often rely on external AI assistance repeatedly in key battle areas or local life-and-death problems during the game, rather than seeking help evenly throughout the board. As a result, spatial clustering of moves with high unattributed components is formed in local areas of the board.

[0079] S433: When a continuously clustered area is detected and the number of pieces placed in the area exceeds the preset clustering threshold, a high-risk cheating determination is directly triggered, and the high-risk cheating determination is output as the result of the cheating risk level.

[0080] A preset clustering threshold is used to define the boundary between sporadic spatial proximity moves and systematic local clustering cheating. When the number of high-unattributed moves within a continuous clustering area exceeds this threshold, it indicates that the candidate has consistently made high-consistency moves in a local area of ​​the chessboard that cannot be explained by visual behavior, forming a chain of evidence of cheating behavior in the spatial dimension. At this time, the system directly triggers a high-risk cheating judgment and outputs this high-risk cheating judgment as the result of the cheating risk level. This judgment path is independent of the threshold comparison and cumulative analysis of the consistency of individual move attributions. Even if the consistency of individual move attributions has not yet reached the trigger threshold, or the cumulative unattributed index has not yet exceeded the overall threshold, as long as a cluster of high-unattributed moves that meet the clustering conditions is detected in the spatial dimension, a high-risk judgment can be made directly.

[0081] The aforementioned non-attributable component spatial clustering detection mechanism extends anti-cheating analysis from the temporal dimension to the spatial dimension. By capturing the local clustering characteristics of cheating behavior on the chessboard, it provides the anti-cheating system with a spatial evidence judgment path independent of temporal cumulative analysis, enabling the effective identification of hidden cheating patterns that repeatedly seek help in key local areas, and further improving the coverage and judgment capability of the multi-dimensional anti-cheating system.

[0082] Furthermore, in step S42, the unattributed component is used as the behavioral attribution consistency score for subsequent cheating risk assessment. After removing the autonomous attribution component from the behavioral attribution consistency score calculation, ensuring that the high consistency score explained by visual behavior does not raise suspicion of cheating, the following sub-steps are also included: S441: Obtain the candidate's personal vision-strength baseline constructed from autonomous attribution components in historical games. The personal vision-strength baseline reflects the candidate's autonomous attribution component level corresponding to different positional complexities under normal thinking conditions.

[0083] The autonomous attribution component is the part of the original AI consistency explained by the candidate's visual attention behavior, reflecting the candidate's ability to make choices consistent with the AI's recommendations under normal observation and thinking conditions. The autonomous attribution component data accumulated by the same candidate in multiple normal games constitutes the correspondence between the candidate's visual behavior and chess strength performance under honest playing conditions. The personal visual-chess strength baseline is a quantitative model of this correspondence. It is constructed by collecting the autonomous attribution component of each move made by the candidate in historical normal games and the corresponding position complexity at the time of the move. Statistics are then performed segmented according to position complexity, and the typical level of the autonomous attribution component within each complexity interval is calculated as the baseline value corresponding to that complexity. Position complexity can be calculated synchronously by the AI ​​engine during the game, for example, by comprehensively determining indicators such as the policy network distribution entropy of candidate moves or the fluctuation range of win rate. Higher complexity indicates more possible branches in the position and greater decision-making difficulty. The personal visual-chess strength baseline reflects the degree to which the candidate's visual attention behavior supports move decisions under different decision-making difficulties during normal thinking, and serves as a personalized reference standard for the candidate's individual behavioral characteristics. The baseline data can be continuously updated as candidates participate in grading exams or play more games, so that the baseline value gradually approaches the candidate's actual behavioral habits.

[0084] S442: Compare the autonomous attribution component of the current move with the baseline value corresponding to the complexity of the current position in the personal vision-strength baseline, and calculate the deviation.

[0085] After semantic decomposition of the current move's consistency, the autonomous attribution component value is obtained. Simultaneously, the complexity score of the situation at the moment of the move is obtained from the AI ​​engine, and the corresponding baseline value is found in the individual's vision-playing ability baseline based on this complexity score. Deviation is the degree of deviation of the current autonomous attribution component from the baseline value, which can be determined by calculating the difference or ratio between the current value and the baseline value. When the deviation is small, it indicates that the candidate's current move's autonomous attribution component is basically consistent with their historical normal level, and the degree to which visual attention behavior supports the move decision is consistent with their personal behavioral habits. When the deviation is large, it indicates that the candidate's current move's autonomous attribution component deviates significantly from their historical normal level. It is worth noting that deviation can occur in two directions: a significantly lower autonomous attribution component than the baseline value may suggest that the candidate made a high-consistency move without conducting normal level visual observation; a significantly higher autonomous attribution component than the baseline value may also constitute an abnormal behavior signal, suggesting that the candidate may be deliberately changing their vision behavior pattern to avoid detection.

[0086] S443: When the deviation exceeds the preset deviation threshold, increase the cheating risk level of the current move, even if the unattributed component of the current move has not reached the risk trigger threshold.

[0087] In existing cheating risk assessment systems, unattributed components are the primary indicator triggering risk alarms; when the unattributed component exceeds a certain threshold, the corresponding move is marked as suspicious. However, in some cheating scenarios, cheaters may manipulate the value of the unattributed component by alternating between high-match and low-match moves, keeping it consistently below the trigger threshold to evade detection. The personal line-of-sight-baseline deviation mechanism introduced in this invention provides a supplementary assessment path for these scenarios: when the self-attributed component of a move deviates significantly from the candidate's historical normal level, even if the unattributed component value itself is not prominent and has not yet reached the risk trigger threshold, the system can still increase its cheating risk level based on the magnitude of the deviation. The core logic of this mechanism is that cheating behavior is not only reflected in the absolute value of "matches that cannot be explained by line of sight," but may also be reflected in "abnormal changes in the candidate's behavioral patterns relative to their normal state." The determination of deviation can be independent of the absolute value of the unattributed component; the two complement each other, forming a more comprehensive risk assessment system.

[0088] The aforementioned individual vision-chess skill baseline deviation detection mechanism extends the anti-cheating judgment criteria from a fixed threshold universally applied to a personalized behavioral baseline deviation analysis. This enables the system to identify hidden anomalies that are "inconsistent with the candidate's normal behavioral habits," effectively compensating for the limitations of universal thresholds in dealing with individual differences among candidates and anti-detection strategies. This further improves the judgment accuracy and robustness of the anti-cheating system.

[0089] Furthermore, in step S42, the unattributed component is used as the behavioral attribution consistency score for subsequent cheating risk assessment. After removing the autonomous attribution component from the behavioral attribution consistency score calculation, ensuring that the high consistency score explained by visual behavior does not raise suspicion of cheating, the following sub-steps are also included: S451: For a continuous sequence of moves, detect whether there is a periodic alternation pattern where line-of-sight deviation and high-match moves occur alternately. Line-of-sight deviation is defined as a line-of-sight matching coefficient below a preset deviation threshold, and high-match moves are defined as moves whose unattributed components exceed a preset high-match threshold. A periodic alternation pattern is defined as a pattern in which line-of-sight deviation and high-match moves alternate over a period of time, and the number of alternations exceeds a preset alternation number threshold.

[0090] "Gaze deviation" refers to a state where the gaze-placement matching coefficient is lower than a preset deviation threshold. This indicates that the examinee's visual attention was significantly insufficient towards the placement area within a preset time window before the move, and the gaze may have strayed from the board or failed to effectively focus on the placement area. "High-matching move" refers to a move where the unattributed component exceeds a preset high-matching threshold. This indicates that a high proportion of the original AI match rate for the move cannot be explained by gaze behavior; that is, the examinee made a move highly consistent with the AI ​​recommendation without effective visual attention support. "Periodic alternation pattern" refers to the alternation of the above two states in a sequence of moves over a period of time, forming a repetitive rhythm of "gaze deviation—high-matching move—gaze deviation—high-matching move," and the number of such alternations exceeds a preset alternation count threshold. The alternation count is calculated as follows: in a continuous sequence of moves, the transition between the gaze deviation state and the high-matching move state is identified, and each transition from one state to the other is counted as one alternation. When the cumulative number of alternations within a preset observation window reaches the preset alternation count threshold, a periodic alternation pattern is determined to exist. This periodic alternation pattern is highly consistent with the rhythm of typical decentralized cheating: the cheater obtains move suggestions from off-board sources when their gaze is off the board, then makes a highly relevant move before their gaze has fully returned to the board, and then shifts their gaze again to obtain the next move suggestion, repeating this cycle. Since each instance of cheating is only manifested in a single move, and both the gaze deviation and the high relevance can be controlled at threshold levels, a single event is difficult to independently determine as cheating. However, its periodic repetition provides a unique basis for identifying this type of behavior.

[0091] S452: When a periodic alternation pattern is detected, a high-risk cheating determination is directly triggered, and the high-risk cheating determination is output as the result of the cheating risk level, instead of relying on the attribution consistency of a single move.

[0092] This judgment path is independent of the attribution consistency threshold for individual moves. It no longer accumulates evaluations based on the isolated risk scores of each move, but instead treats the periodic alternation pattern itself as a holistic, highly credible piece of evidence of cheating. The basic logic of this design is that the rhythmic alternation of gaze deviation and highly consistent moves has almost no reasonable cognitive or behavioral explanation in normal gameplay. During normal gameplay, a player's gaze level and move quality do not exhibit regular alternating fluctuations; instead, this rhythm highly matches the behavior pattern of intermittently consulting external cues. Therefore, once a periodic alternation pattern that meets the alternation frequency threshold is detected, a high-risk cheating judgment can be made directly, without needing to analyze whether the attribution consistency reaches the risk trigger threshold for each move.

[0093] The aforementioned periodic alternation pattern detection mechanism extends anti-cheating analysis from single-point judgment, trend analysis, and spatial clustering to the level of behavioral rhythm recognition, providing an effective means to capture dispersed, intermittent, and highly anti-detection-conscious covert cheating behaviors. This mechanism, together with semantic segmentation based on the consistency of single-hand moves, temporal deviation detection based on consecutive moves, spatial clustering analysis based on the entire game, and deviation detection based on individual historical baselines, constitutes a multi-layered, multi-perspective anti-cheating judgment system, significantly improving the system's ability to identify and cover various cheating behavior patterns and its robustness in judgment.

[0094] The above method will be further illustrated below with several embodiments: Example 1 This embodiment uses an online Go ranking tournament as an application scenario to describe the complete execution flow of the method of the present invention in detail. The Go interaction device uses an embedded main control chip as the local processing unit with a main frequency of 1.8GHz, and integrates a neural network processor for real-time inference of the local visual model. The first image acquisition module is an RGB camera installed on the top of the device's pillar, with a resolution of 1920×1080, a frame rate of 30 frames / second, a horizontal field of view of 120 degrees, and is installed at a downward angle towards the examinee's face, at a height of about 45 cm from the table, which can completely cover the examinee's head and upper body area. The second image acquisition module is a top-down camera installed on the bracket directly above the Go board, with a resolution of 2592×1944, a frame rate of 15 frames / second, and an optical axis perpendicular to the plane of the Go board, which can completely cover all intersections of a standard 19×19 Go board. The device is equipped with a Wi-Fi 5 communication module, which transmits data with the Go AI engine deployed in the cloud through an encrypted network link, and the round-trip latency of a single AI decision request is controlled within 2 seconds.

[0095] The chessboard plane is pre-divided into 36 6×6 pre-defined areas, each covering approximately 3×3 to 4×4 chessboard intersections. The boundaries of these pre-defined areas are stored in the non-volatile memory of the local processing unit. The rigid transformation matrix between the optical center of the first image acquisition module and the chessboard world coordinate system is calibrated using the Zhang Zhengyou calibration method before the device leaves the factory. The calibration parameters are written into the device firmware and automatically loaded upon power-on. Candidates must complete identity registration upon first use of the device. During registration, the system acquires the candidate's facial image and extracts facial feature vectors. These feature vectors are then bound to the candidate's identity information and stored in the cloud-based candidate database.

[0096] Before the exam begins, the system enters the identity verification phase. The first image acquisition module captures the candidate's facial image, the local processing unit runs a lightweight face detection model to locate the facial region, extract facial feature vectors, and sends the feature vectors to the cloud via the network module for a one-to-one comparison with registered features. If the comparison score exceeds the preset identity verification threshold of 0.85, the identity verification is successful, the system unlocks the exam function and enters the game state. If the identity verification fails, the system locks the operation interface and sends an identity anomaly notification to the referee management backend via the network module.

[0097] After the game begins, the first image acquisition module continuously captures facial images of the examinee at a frame rate of 30 frames per second. The local processing unit runs a facial feature point detection model on each frame, extracting the coordinates of 68 facial feature points, including the center points of the eyes, the tip of the nose, and the corners of the mouth. Based on the feature point coordinates, the PnP algorithm is used to calculate the three-dimensional rotation angle of the head, including pitch, yaw, and roll angles. Using the direction of the head towards the normal as an approximation of the gaze direction vector, and combining it with the pre-calibrated camera extrinsic matrix, the intersection of the gaze direction ray and the chessboard plane is solved to obtain the gaze chessboard coordinates corresponding to that frame. The preset area number of the chessboard where the gaze chessboard coordinates fall is the gaze area label for that frame. The gaze area labels of all frames are arranged in chronological order to form a gaze chessboard area sequence. Simultaneously, the second image acquisition module continuously captures chessboard images at a frame rate of 15 frames per second. The local processing unit runs a piece detection model, identifies the position of newly added pieces based on the difference image between the current frame and the previous frame, converts the piece image coordinates into chessboard world coordinates through a perspective transformation matrix, and records the placement coordinates and placement time.

[0098] The following example illustrates the subsequent processing flow using a specific move made during the game. The move occurred at 18 minutes, 32 seconds, and 500 milliseconds, with the move coordinates (12, 8) on the chessboard and the move area being the upper right corner of the board. A preset time window length of 3 seconds was set, and the gaze sequence of the chessboard area within the 3 seconds prior to the move was extracted, totaling 90 frames of gaze data. Statistical analysis revealed 82 valid gaze frames out of these 90 frames. Of these, 51 frames matched the gaze area with the move's coordinates, resulting in a matching ratio of approximately 0.62 (51 divided by 82). The percentage of valid gaze frames (82 divided by 90) is approximately 0.91, exceeding the preset threshold of 0.70, indicating reliable gaze data. The gaze-move matching coefficient is then used in the subsequent calculation process.

[0099] In the time-series segmentation phase, the system calculates the gaze distribution entropy value at each sampling moment within a preset time window, with a sampling interval of 0.2 seconds, totaling 15 sampling points. The gaze distribution entropy values ​​at the first 8 sampling points are all higher than the preset entropy threshold of 1.5. Starting from the 9th sampling point, the entropy value continuously decreases, falling below the preset entropy threshold by the 12th sampling point. The system uses the moment corresponding to the 10th sampling point as the first boundary between the scanning phase and the focusing phase. The scanning phase lasts approximately 2.0 seconds, and the focusing phase lasts approximately 1.0 second. During the scanning phase, the system counts 14 changes in the gaze area, covering 11 different preset chessboard areas, for a total of 36 preset chessboard areas. The saccade transfer rate is calculated as 14 divided by 2.0, which equals 7.0 times / second. The chessboard coverage rate is calculated as 11 divided by 36, which equals approximately 0.31. The saccade path complexity is the product of these two values, which is 7.0 multiplied by 0.31, which equals 2.17. During the focusing phase, the system divides the focusing phase into a first half and a second half. The frequency of occurrence of the area where the piece coordinates are located in the first half is 0.35, and the frequency in the second half is 0.72. The frequency in the second half is greater than that in the first half and exceeds the preset focusing frequency threshold of 0.60, indicating that a focusing convergence process exists. In the transition pattern recognition stage, the system detects that in the four frames before entering the area where the piece coordinates are located, the gaze area gradually transitions from the lower left corner of the chessboard, which is farther away from the piece area, through the adjacent area in the middle of the chessboard, to the piece area, indicating that the transition pattern is a progressive convergence pattern. Based on the above analysis results, the scan path complexity of 2.17 is higher than the preset complexity threshold of 1.50, which meets the preset exploration requirements, and a focusing convergence process exists. The transition pattern is a progressive convergence pattern, and the system generates a line-of-sight matching coefficient of 0.82 for this move.

[0100] Simultaneously, the system sends the complete game record up to the current move to the cloud AI engine via an encrypted network link. The cloud AI engine loads the current game state, runs the policy network and value network to perform a Monte Carlo tree search, and outputs the probability distribution of the AI's preferred move and the top 10 candidate moves. The current move is completely consistent with the AI's preferred move, and the original AI's consistency P is achieved. raw The system uses a line-of-sight matching coefficient of 0.82 to semantically decompose the original AI consistency score. The self-attributed component is calculated as 1.0 multiplied by 0.82, which equals 0.82, and the unattributed component is calculated as 1.0 minus 0.82, which equals 0.18. The unattributed component of 0.18 is output as the behavioral attribution consistency score to the risk assessment module. If the behavioral attribution consistency score of 0.18 is lower than the preset high-risk threshold of 0.55 and lower than the preset medium-risk threshold of 0.35, it falls into the low-risk range. The system marks the move as low-risk, records the behavioral data, and does not trigger an alarm.

[0101] Throughout the later stages of the game, the candidate's behavioral attribution consistency remained between 0.15 and 0.35, without triggering any medium- or high-risk alerts. The cumulative unattributed index for the entire game was 12.3, which is lower than the preset cumulative threshold of 20.0. No continuous clustering areas were detected in the unattributed spatial distribution map, and no line-of-sight-consistency deviation anomalies were detected in the joint time series analysis. The system determined that the overall risk level of the game was low, and the exam score was valid.

[0102] Example 2 This embodiment uses the scenario of a candidate intermittently cheating during certain stages of an online Go ranking exam using an external AI tool to demonstrate the identification process of dynamic cheating behavior by the method of this invention. The device hardware configuration and parameter settings are the same as in Embodiment 1.

[0103] In the first 60 moves of the exam game, the candidate's visual attention behavior was normal: the average saccadic path complexity remained between 1.80 and 2.50, the vast majority of moves showed a focus convergence process, the eye movement into the move area was mainly a gradual convergence pattern, the autoattribution component was generally in the range of 0.65 to 0.88, the unattributed component was generally in the range of 0.05 to 0.25, and the behavioral attribution consistency remained in the low-risk range. The cumulative unattributed index for the 26 moves from move 35 to move 60 was only 4.6.

[0104] Starting from move 61, the candidate's behavior pattern changed significantly. The system constructed a joint time series for moves 61 to 80 and applied a sliding detection window with a window length of 8 moves. The average original AI consistency within the preceding window (moves 53 to 60) was 0.35. Within the detection window (moves 61 to 68), the average original AI consistency jumped to 0.72, an increase of 0.37, exceeding the preset increase threshold of 0.25. Simultaneously, the average line-of-sight matching coefficient within the same detection window was 0.24, a change of only 0.03 compared to the average line-of-sight matching coefficient of 0.68 in the preceding window, below the preset change threshold of 0.15. With a significant increase in original AI consistency while the line-of-sight matching coefficient remained essentially unchanged, the system determined that the move sequence from move 61 to 68 exhibited an abnormal discrepancy in line-of-sight consistency.

[0105] Further analysis of the specific data for each move within the window revealed the following: Before the 61st move, the candidate's gaze deviated from the chessboard for 0.65% of the time within the preset time window, and the proportion of effective gaze frames matching the move's location was only 0.22, resulting in a gaze-move matching coefficient of 0.18. Simultaneously, this move perfectly matched the AI's preferred placement, with an original AI consistency of 1.0. After semantic segmentation, the autonomous attribution component was 0.18, and the unattributed component was 0.82. Moves 63 and 65 also exhibited a gaze-move matching coefficient below 0.30 while the original AI consistency was above 0.85. The system applied a preset deviation penalty factor of 1.4 to each of the eight moves within the deviation anomaly window. The attribution consistency of each move, multiplied by 1.4 from the original unattributed component, generally exceeded 0.55, entering a high-risk range. Meanwhile, in the sequence from move 61 to move 68, the system detected four instances of alternating between a line-of-sight deviation and a high-matching move, exceeding the preset threshold of three alternations. This indicated a periodic alternation pattern, directly triggering a high-risk cheating determination. The system automatically pushed an alarm notification to the referee management backend, along with synchronized video clips from the first and second image acquisition modules during the alarm period, as well as an unattributed spatial distribution map, for the referees to review manually.

[0106] Example 3 This example uses a high-level junior Go player's normal game in an online Go ranking tournament as a scenario to demonstrate how to effectively distinguish between a high degree of consistency generated by independent observation and thinking and an abnormally high degree of consistency lacking visual support, thus avoiding false alarms about high-level players.

[0107] The candidate is a registered amateur 5-dan player with the Chinese Weiqi Association, possessing 7 years of experience. Through daily training and competitions, they have accumulated over 300 valid game records, and their personal vision-strength baseline has been constructed in the system. The baseline construction process is as follows: the system extracts the autonomic attribution components and corresponding positional complexity of each move in the candidate's historical games. The positional complexity is divided into 5 intervals with a step size of 2.0, ranging from 0 to 10. The median of the autonomic attribution components of all moves within each interval is taken as the baseline value for that complexity interval. The candidate's baseline value for the autonomic attribution components is 0.72 in the complexity interval of 2.0 to 4.0, 0.65 in the complexity interval of 4.0 to 6.0, 0.55 in the complexity interval of 6.0 to 8.0, and 0.42 in the complexity interval of 8.0 to 10.0. Baseline data indicates that the candidate possesses strong independent thinking and situational judgment abilities. Under normal playing conditions, the candidate's visual attention behavior during most moves effectively supports their choice of moves consistent with AI recommendations.

[0108] In the 78th move of this professional ranking tournament, the game entered a complex mid-game battle phase, with the AI ​​engine assessing the current situation's complexity at 7.2. Before this move, the system detected a scan path complexity of 2.85 within the preset time window, significantly higher than the preset complexity threshold of 1.50. During the focusing phase, the frequency of the move's placement area increased from 0.28 in the first half to 0.78 in the second half, indicating a focusing convergence process. The eye movement into the placement area followed a gradual convergence pattern, transitioning through three adjacent preset areas on the board. Considering all indicators, the eye-placement matching coefficient for this move was 0.91. This move aligned with the preferred placement point of the cloud-based AI engine, resulting in an initial AI consistency of 1.0. Semantic analysis revealed that the autonomous attribution component was 1.0 multiplied by 0.91, equaling 0.91, while the non-attribution component was 1.0 minus 0.91, equaling 0.09. The behavioral attribution consistency was 0.09, placing it in the low-risk range.

[0109] At this point, if the existing AI consistency analysis scheme, which does not distinguish between visual attention behavior, is used, the consistency of this move would be 1.0, and it would be directly marked as high risk or at least medium risk. However, the method of this invention, through analysis of the visual attention behavior data, finds that: before the move, the examinee underwent a thorough global scan (scanning path complexity 2.85), possessed a complete focusing and convergence process, and entered the move area with a progressive convergence mode, forming a complete "observation-focusing-move" behavior chain. The visual attention behavior provided sufficient attribution support for the high consistency of this move. Therefore, most of the consistency was classified into the autonomous attribution component and eliminated, with only the 0.09 that could not be attributed being considered as a valid risk input. Simultaneously, the system compares the autonomous attribution component of this move (0.91) with the baseline value of 0.55 in the candidate's personal vision-chess strength baseline corresponding to the current complexity of 7.2 (complexity 6.0 to 8.0). The deviation is 0.91 minus 0.55, which equals 0.36, exceeding the preset deviation threshold of 0.25, indicating a positive deviation. In this context, a positive deviation reflects that the candidate's visual attention and autonomous thinking quality in the current complex situation are significantly higher than their historical normal levels. Combined with the high scan path complexity and clear focusing convergence characteristics of this move, the system determines that the deviation is due to the candidate concentrating on in-depth thinking when facing a critical battle situation, maintaining a low-risk judgment and not triggering a false alarm.

[0110] After the entire game, the candidate's attribution consistency remained between 0.03 and 0.22 throughout, with a cumulative unattributed index of 8.7. No local clustering was observed in the unattributed spatial distribution map. No line-of-sight divergence anomalies were detected in the joint time series analysis, and no periodic alternation patterns were detected in the line-of-sight alternation detection. The system comprehensively determined the game's risk level to be low, and the exam result was valid. This embodiment demonstrates that the method of the present invention can effectively distinguish between the normal high-level performance of a high-level player and cheating behavior, fundamentally solving the technical defect of existing AI consistency analysis schemes that are prone to false positives against strong players.

[0111] Accordingly, a second aspect of the present invention provides an electronic device, including: at least one processor and a memory connected to the at least one processor. The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, cause the at least one processor to perform the visual detection anti-cheating method of the aforementioned chess-playing interactive device.

[0112] Accordingly, a third aspect of the present invention provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the visual detection anti-cheating method of the above-described chess-playing interactive device.

[0113] The embodiments of the present invention aim to protect a visual detection anti-cheating method for a chess-playing interactive device, which has the following effects: 1. By decomposing the candidate's gaze behavior before placing a piece into a scanning phase and a focusing phase, and extracting the scanning path complexity and focusing convergence features, this invention can restore the complete visual thinking chain of "global observation - local focusing - confirmation of placement". This improves the system's evaluation of the candidate's gaze behavior from the static duration statistics of the existing technology to dynamic temporal structure analysis, thereby effectively distinguishing between the two behavior modes of "placing a piece after sufficient observation and consideration" and "placing a piece directly without observing the chessboard", which significantly improves the accuracy of cheating behavior identification and the interpretability of the judgment results. 2. By splitting the original AI consistency score into an autonomous attribution component and an unattribution component, and removing the autonomous attribution component from the cheating risk calculation, this invention enables the consistency score index to carry behavioral attribution information. The high consistency score generated by high-level players under normal thinking conditions is automatically classified into the autonomous attribution component and is exempt from triggering alarms. Only the unattribution component, which lacks visual support, is used as an effective risk input. This fundamentally solves the inherent defect in existing AI consistency score analysis schemes where strong players are easily falsely accused, and significantly reduces the false alarm rate. 3. By further introducing joint time-series deviation detection, chessboard space clustering analysis, individual line-of-sight-chess strength baseline deviation detection, and line-of-sight deviation-high-match periodic alternation pattern detection on the basis of semantic segmentation of consistency, this invention constructs a complete system for three-dimensional identification of cheating behavior from multiple dimensions of time series, space, individual baseline, and behavioral rhythm. This enables decentralized cheating behaviors with high concealment and weak single-instance evidence to be locked by multi-dimensional cross-validation, significantly improving the identification coverage and judgment robustness of complex cheating methods in online examination scenarios.

[0114] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0115] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0116] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0117] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A visual detection method for preventing cheating in a chess-playing interactive device, characterized in that, Includes the following steps: S1: The candidate's facial image is acquired by the first image acquisition module pointing towards the candidate's face, the candidate's gaze area is calculated in real time, and the gaze area is mapped to a pre-calibrated chessboard coordinate system to obtain a gaze chessboard region sequence; wherein, the gaze chessboard region is the region in which the mapping point of the gaze area falls on the chessboard plane after the chessboard plane is divided into multiple preset regions. S2: Detect the move event by using the second image acquisition module pointing to the chessboard area, obtain the coordinates of the current move in the chessboard coordinate system, and record the move time; S3: For the current move, extract the gaze chessboard region sequence within a preset time window before the move. Calculate the gaze-move matching coefficient based on the proportion of valid gaze frames matching the region where the move is located to the total number of valid gaze frames within the preset time window. When the proportion of valid gaze frames within the preset time window is lower than a preset threshold, mark the gaze-move matching coefficient as invalid. S4: Obtain the original AI consistency calculated by the AI ​​engine for the current move, and use the line-of-sight-move matching coefficient to correct the original AI consistency and generate the behavior attribution consistency; when the line-of-sight-move matching coefficient is invalid, the original AI consistency is directly used as the behavior attribution consistency. S5: Based on the attribution consistency of the behavior and the preset risk classification strategy, determine the cheating risk level of the current move and the game.

2. The visual detection anti-cheating method for the chess-playing interactive device according to claim 1, characterized in that, The calculation of the line-of-sight-piece matching coefficient in step S3 includes the following sub-steps: S31: Divide the preset time window into at least two timing stages, the at least two timing stages including a scanning stage and a focusing stage; S32: During the scanning phase, the scanning path complexity is calculated based on the transition characteristics of the gaze area in the gaze area sequence between each preset area of ​​the chessboard. The scanning path complexity is used to measure the degree of global exploration of the candidate's gaze on the chessboard. S33: During the focusing phase, based on the frequency change characteristics of the area where the move coordinates are located in the sequence of the gaze chessboard areas, it is determined whether there is a focusing convergence process. The focusing convergence process is the process in which the examinee's gaze gradually focuses on the area where the move coordinates are located as the move time approaches. S34: Generate the line-of-sight matching coefficient based on the complexity of the scanning path and the presence or absence of the focusing convergence process; wherein, when the complexity of the scanning path does not meet the preset exploration requirements, the line-of-sight matching coefficient is reduced; when the focusing convergence process exists, the line-of-sight matching coefficient is increased.

3. The visual detection anti-cheating method for the chess-playing interactive device according to claim 2, characterized in that, Before generating the line-of-sight-piece matching coefficient in step S34, the following sub-steps are also included: S3401: During the focusing phase, identify the transition pattern of the gaze entering the area where the piece is placed, the transition pattern including at least a skip-direct approach mode and a progressive convergence mode; wherein, the skip-direct approach mode is that in the frame before entering the area where the piece is placed, the gaze area is not adjacent to the area where the piece is placed; the progressive convergence mode is that in the several frames before entering the area where the piece is placed, the gaze area gradually transitions from an area far from the piece is placed to the area where the piece is placed through adjacent areas. S3402: When the identified transfer mode is the skip-direct mode, a preset transfer mode penalty factor is applied to the line-of-sight matching coefficient to reduce the line-of-sight matching coefficient; when the identified transfer mode is the progressive convergence mode, the line-of-sight matching coefficient is kept unchanged.

4. The visual detection anti-cheating method for the chess-playing interactive device according to claim 2, characterized in that, Step S31, which involves dividing the preset time window into at least two time-series stages, includes the following sub-steps: S311: Calculate the gaze distribution entropy value of the gaze chessboard region sequence at each moment within the preset time window. The gaze distribution entropy value is the information entropy of the probability distribution of each preset region being gazed upon. The higher the gaze distribution entropy value, the more uniform the distribution of the examinee's gaze among the preset regions of the chessboard. S312: Detect the turning point in time when the gaze distribution entropy value continuously decreases from above a preset entropy threshold to below the preset entropy threshold; S313: The detected turning point is used as the first boundary between the scanning phase and the focusing phase, instead of the preset fixed boundary.

5. The visual detection anti-cheating method for the chess-playing interactive device according to claim 2, characterized in that, After generating the line-of-sight matching coefficients based on the scan path complexity and the presence or absence of the focusing convergence process in step S34, the following sub-steps are also included: S351: When the complexity of the scanning path is lower than a preset complexity threshold and the focusing convergence process is not detected, the current move is marked as a visually disjointed move. S352: When the number of consecutive visually disjointed moves exceeds a preset disjoint sequence length threshold, a high-risk cheating determination is directly generated, and the high-risk cheating determination is output as the result of the cheating risk level, no longer relying on the line-of-sight matching coefficient and the behavior attribution consistency.

6. The visual detection anti-cheating method for the chess-playing interactive device according to claim 1, characterized in that, Step S4, which involves using the line-of-sight matching coefficient to correct the original AI fit and generate a behavior attribution fit, includes the following sub-steps: S41: Based on the line-of-sight matching coefficient, the original AI matching degree is split into the part that can be attributed to the candidate's autonomous visual attention and the part that cannot be attributed to the candidate's autonomous visual attention, and autonomous attribution component and non-attribution component are obtained respectively. S42: The unattributed component is used as the behavioral attribution consistency for subsequent cheating risk determination; wherein, the autonomous attribution component is removed from the calculation of the behavioral attribution consistency, so that the high consistency part explained by the visual behavior does not generate cheating suspicion.

7. The visual detection anti-cheating method for the chess-playing interactive device according to claim 6, characterized in that, Step S41, which involves splitting the original AI matching degree into a part attributable to the examinee's autonomous visual attention and a part not attributable to the examinee's autonomous visual attention based on the gaze-to-placement matching coefficient, and obtaining the autonomous attribution component and the non-attribution component respectively, further includes the following sub-steps: S411: For consecutive multiple moves, construct a joint time series consisting of the line-of-sight-move matching coefficient sequence and the original AI consistency sequence; S412: Slide the detection window on the joint time series. When the increase of the mean value of the original AI matching degree in the detection window compared with the previous window exceeds the preset increase threshold, and the change of the mean value of the line-to-spot matching coefficient in the same detection window is lower than the preset change threshold, it is determined that there is a line-to-spot matching degree deviation anomaly. S413: When it is determined that there is an abnormality in the line-of-sight consistency, a preset deviation penalty factor is applied to the behavior attribution consistency of each piece in the detection window to increase the cheating risk level of the corresponding piece.

8. The visual detection anti-cheating method for the chess-playing interactive device according to claim 6, characterized in that, Step S42, which involves using the unattributed component as the behavioral attribution consistency for subsequent cheating risk determination, further includes the following sub-steps after the autonomous attribution component is removed from the calculation of the behavioral attribution consistency, ensuring that the high consistency portion explained by visual behavior does not generate cheating suspicion: S431: Map the unattributed components of each move in the entire game to the chessboard coordinate system according to their move coordinates, and construct an unattributed spatial distribution map. S432: On the unattributed spatial distribution map, detect whether there are regions on the chessboard where pieces with unattributed components exceeding a preset high value threshold are continuously clustered; the continuously clustered region is a connected region on the chessboard where pieces with high unattributed components exist in adjacent preset regions. S433: When the continuously clustered area is detected and the number of pieces placed in the area exceeds the preset clustering number threshold, a high-risk cheating determination is directly triggered, and the high-risk cheating determination is output as the result of the cheating risk level.

9. The visual detection anti-cheating method for the chess-playing interactive device according to claim 6, characterized in that, Step S42, which involves using the unattributed component as the behavioral attribution consistency for subsequent cheating risk determination, further includes the following sub-steps after the autonomous attribution component is removed from the calculation of the behavioral attribution consistency, ensuring that the high consistency portion explained by visual behavior does not generate cheating suspicion: S441: Obtain the candidate’s personal vision-chess strength baseline constructed by the autonomous attribution components in historical games, wherein the personal vision-chess strength baseline reflects the level of autonomous attribution components corresponding to different board complexities under normal thinking conditions. S442: Compare the autonomous attribution component of the current move with the baseline value corresponding to the current position complexity in the personal vision-chess strength baseline, and calculate the deviation. S443: When the deviation exceeds the preset deviation threshold, the cheating risk level of the current move is increased, even if the unattributed component of the current move has not reached the risk trigger threshold.

10. The visual detection anti-cheating method for the chess-playing interactive device according to claim 6, characterized in that, Step S42, which involves using the unattributed component as the behavioral attribution consistency for subsequent cheating risk determination, further includes the following sub-steps after the autonomous attribution component is removed from the calculation of the behavioral attribution consistency, ensuring that the high consistency portion explained by visual behavior does not generate cheating suspicion: S451: For a continuous sequence of moves, detect whether there is a periodic alternation pattern in which line-of-sight deviation and high-match moves alternate; the line-of-sight deviation is the state in which the line-of-sight-move matching coefficient is lower than a preset deviation threshold, and the high-match moves are moves in which the unattributed component exceeds a preset high-match threshold; the periodic alternation pattern is a pattern in which the line-of-sight deviation and the high-match moves alternate within a period of time and the number of alternations exceeds a preset number of alternations threshold. S452: When the periodic alternation pattern is detected, a high-risk cheating determination is directly triggered, and the high-risk cheating determination is output as the result of the cheating risk level, instead of relying on the behavior attribution consistency of a single move.

Images