Artificial intelligence base system for secure and fair online reviews

Abstract

An artificial intelligence-based system for safe and fair online reviews. The system includes: an investigation processing device comprising at least one high-resolution camera, a directional microphone array, a secure processor and a non-volatile storage medium; a facial recognition module running on the secure processor and configured to continuously authenticate a test subject by comparing live facial features with pre-stored biometric templates; an eye-tracking module that determines the angular displacement of the subject's head and eyes in relation to the display screen and generates deviation signals when thresholds are exceeded; a control module for object recognition that is configured to analyze video images captured by the camera in order to identify unauthorized devices, printed materials, or other people in the field of view; an audio anomaly detection module that processes input signals from the microphone array and detects unauthorized speech activity, secondary human voices, or irregular background noise patterns; a probability-based misbehavior detection engine that calculates a probability score for misbehavior by assigning weighted coefficients to the outputs of the face detection module, the eye tracking module, the object detection control module, and the audio anomaly detection module; and an automated scoring engine configured to evaluate both objective and subjective responses from candidates, with subjective responses being processed using semantic and syntactic natural language processing models to generate a score, and with all results and alerts being transmitted to a remote examiner dashboard via an encrypted communication channel.

Description

Technical field

[0001] The invention relates to the field of online testing systems, digital authentication devices, and automated monitoring technologies. More specifically, it is a device-enhanced software-hardware platform that utilizes artificial intelligence, computer vision, and multimodal sensors for secure, scalable, and fair online assessments. Background of the invention

[0002] With the increasing prevalence of online education, remote examinations have become indispensable. Traditional online tests often lack integrity due to identity theft, unauthorized device use, and a lack of real-time monitoring. Existing proctoring solutions rely primarily on webcams and rudimentary surveillance, which are insufficient to detect sophisticated cheating attempts. Furthermore, many systems lack a hardware-based security layer, allowing for circumvention through software manipulation. Therefore, there is an urgent need for an integrated platform that combines hardware, software, and AI-driven detection to ensure secure and fair online examinations.

[0003] The development of online education and remote examinations has presented both opportunities and challenges regarding examination integrity. The technological background of online proctoring stems from the gradual shift from traditional face-to-face examinations to web-based testing platforms, driven by global digital transformation, the expansion of internet infrastructure, and the widespread availability of personal computers. Online examinations were initially designed to replicate traditional examination methods in a digital format, offering advantages such as flexibility, reduced logistical effort, and improved accessibility. However, the absence of physical proctors has created vulnerabilities that have allowed students to engage in abuse, including identity theft, unauthorized use of equipment, and collusion with external parties.To address these vulnerabilities, various solutions have been developed over time, each utilizing different technological foundations such as browser-based restrictions, video surveillance, and artificial intelligence. Despite gradual improvements, these approaches continue to suffer from limitations in terms of accuracy, fairness, and robustness against sophisticated fraud attempts.

[0004] One of the first solutions was browser blocking systems, designed to restrict student activity in a secure exam environment. Such systems used tab locking mechanisms, keyboard input restrictions, and application access controls to prevent candidates from switching between resources during the exam. While these methods were effective against casual attempts to search for information or communicate digitally, they had fundamental drawbacks. Tech-savvy candidates could circumvent these restrictions by using external devices such as smartphones, tablets, or additional computers, which remained invisible to the monitoring system. Furthermore, browser blocking systems required a high degree of reliance on software stability and network integrity, frequently resulting in false alarms or unintended interruptions when legitimate system functions were mistakenly flagged as suspicious behavior.This led to frustration among the candidates and reduced confidence in the examination process.

[0005] Another widespread solution was webcam-based video surveillance, in which a candidate's live feed was recorded and either observed by human supervisors or recorded for later review. These systems initially represented a significant improvement, as they introduced a human-in-the-loop approach to ensure fairness. However, the reliance on manual human supervisors was fraught with scalability issues. Monitoring large groups of students across multiple time zones required enormous staffing, and the subjective nature of human observation often led to inconsistencies in reporting suspected misconduct. Furthermore, passive webcam surveillance lacked the granularity to detect more subtle forms of cheating, such as gaze deviations, hidden microphones, or small handheld devices outside the camera's field of view.Because the candidates had to continuously stream their surroundings, privacy concerns arose, leading to ethical debates about the limits of surveillance in academic settings.

[0006] Later advances used computer vision and artificial intelligence to automate surveillance tasks. Researchers introduced facial recognition, facial analysis, and head-posture estimation techniques to confirm candidates' identities and assess their attention. Systems such as those based on Viola Jones facial recognition or local binary pattern descriptors attempted to efficiently detect faces. However, these techniques often struggled with variations in lighting, camera quality, or obstructions from glasses or headwear. Accuracy was compromised in uncontrolled domestic settings where the candidate's situation could not be standardized. Furthermore, many of these techniques were limited to static facial recognition and could not dynamically account for behaviors such as gaze shifts or movements that might indicate misconduct.As a result, students who exploited blind spots in the techniques were still able to use unfair practices.

[0007] Alongside facial recognition, efforts were made to develop speech and audio monitoring systems that recorded microphone input during investigations. These systems aimed to detect unauthorized conversations, the use of voice assistants, or acoustic signals from undercover employees. Signal processing techniques such as fast Fourier transform and spectral feature extraction were employed to distinguish human voices from background noise. While such systems could effectively detect obvious violations, they frequently suffered from a high false-positive rate. Background noise from shared living spaces, ambient sounds, or even echoes were sometimes incorrectly classified as suspicious activity. Furthermore, candidates in noisy or communal environments were unfairly penalized even though they had committed no wrongdoing.These limitations reduced the reliability of pure audio detection and made the combination of multiple sensor modalities necessary for better results.

[0008] Tab activity detection and system event monitoring also became popular features of proctoring solutions. Applications recorded keystrokes, mouse actions, and the frequency of tab switches, generating alerts or automatically terminating the application when certain thresholds were exceeded. While this improved traceability within the software environment, it had the drawback of limited scope. The system could only monitor the machine on which the test was being conducted and was unable to detect misbehavior related to external collaboration or hidden devices. Furthermore, excessive reliance on tab monitoring sometimes penalized legitimate system behavior, such as accidentally minimizing windows or troubleshooting network issues, unfairly disadvantaging honest candidates.

[0009] A growing trend in recent years has been the integration of multimodal recognition, combining facial recognition, eye tracking, object detection, and audio anomaly analysis. This approach, often marketed as AI-powered proctoring, represented a significant advancement. For example, deep learning-based object detection techniques like YOLO and SSD could identify mobile phones or books in the camera's field of view, while convolutional neural networks improved facial recognition accuracy even under varying lighting conditions. Eye-tracking techniques calculated head position angles and pupil coordinates to determine whether candidates were focused on the exam screen. These systems offered real-time alerts and comprehensive monitoring reports, making them attractive to institutions with large-scale exams. However, their limitations remained.The computational effort was high, requiring powerful local hardware or cloud processing, which increased costs. Latency and bandwidth requirements became significant for candidates with poor internet connections, resulting in inconsistent testing experiences. Furthermore, technical fairness remained a concern, as facial recognition models sometimes exhibited demographic biases, leading to a disproportionate number of false negatives or false positives.

[0010] Some solutions also experimented with integrating biometric authentication devices, such as fingerprint or iris scanners, to improve identity verification. While these methods are highly secure in controlled environments, they proved impractical for distributed online assessments due to a lack of standardized hardware among students. The cost and logistical effort required for widespread distribution of such devices limited their adoption. Furthermore, the use of biometric data poses privacy risks, as sensitive information can be vulnerable to breaches or misuse.

[0011] Another challenge for existing surveillance technologies is scalability. While individual modules like facial recognition or tab locking work well in isolation, scaling to thousands of candidates in real time leads to bottlenecks in data processing, network transmission, and storage. Systems relying on human monitoring quickly become unfeasible, while systems based on advanced AI models require infrastructure investments that smaller institutions cannot afford. The question of cost-effective scalability remains unresolved, as most commercial solutions require trade-offs between the depth of surveillance and the number of concurrently supported users.

[0012] Fairness and user-friendliness also prove to be recurring drawbacks of existing solutions. Systems that aggressively report anomalies can unintentionally penalize legitimate behavior, such as adjusting one's seat position, briefly looking away, or dealing with distractions. Conversely, moderate thresholds may fail to detect subtle cheating attempts. Finding the right balance between sensitivity and specificity is often difficult, leading to dissatisfaction among both students and examiners. Furthermore, continuous monitoring causes psychological stress for candidates. Some report increased anxiety, which negatively impacts their performance. Ethical concerns regarding consent, data storage, and privacy rights further complicate the acceptance of existing monitoring platforms.

[0013] A final drawback is the lack of robust scoring systems integrated into proctoring solutions. While monitoring ensures candidates don't cheat, it doesn't guarantee fair assessment. Many platforms still rely on manually grading subjective answers, leading to human bias and delays. Automated scoring systems exist for objective questions, but they struggle with descriptive responses. Early natural language processing techniques based on keyword matching and syntactic feature extraction were often inaccurate and failed to capture semantic nuances in answers. This resulted in discrepancies between machine scores and human expectations, undermining confidence in automated grading.

[0014] Taken together, these drawbacks highlight the shortcomings of existing proctoring technologies. Browser-based methods cannot prevent the use of external devices, webcam-only solutions are not robust enough against sophisticated cheating, audio analysis suffers from environmental variability, and multimodal AI approaches struggle with computational demands, fairness, and scalability. Biometric authentication provides additional security but is impractical for distributed populations, and scoring engines do not achieve human accuracy with complex answers. Data privacy, ethical considerations, and candidate stress further limit the acceptance of current systems. Overall, while technological advancements have improved online proctoring, no existing solution fully meets the requirements for security, fairness, scalability, usability, and cost-effectiveness.This technical background underlines the need for a new integrated system that combines hardware security, multimodal AI detection, real-time alerts, automated evaluation, and ethical safeguards to ensure the integrity of online reviews. Summary of the invention

[0015] The invention provides a device-based online proctoring system. The system consists of a compact console with a high-resolution camera, microphone array, embedded AI accelerator, and secure storage. A proctoring application runs on the console, verifying identity using deep-learning-based facial recognition, monitoring student activity via eye tracking and mobile device detection, and analyzing audio signals for anomalies. The software system detects misconduct in real time using a weighted probability model. A secure server evaluates candidate responses using natural language processing and automatic scoring modules for subjective answers. Real-time alerts are generated for suspicious activity, and session data is encrypted for post-exam review.

[0016] The main objective of the invention is to provide a secure, reliable, and fair mechanism for online examinations via an integrated, device-based system that combines hardware, software, and AI modules into a unified proctoring and grading platform. The invention seeks to overcome the limitations of existing browser-based and webcam-only solutions by integrating advanced recognition mechanisms into a device that ensures continuous monitoring of the candidate's identity, environment, and behavior throughout the examination. A further objective of the invention is to introduce an automated grading engine capable of evaluating both objective and subjective responses with high accuracy and consistency, thereby reducing human bias and accelerating the grading process.Another objective of the invention is the real-time detection of misconduct such as identity theft, gaze deviation, tab switching, unauthorized device use, and background communication. Examiners are immediately notified, and appropriate corrective actions are taken automatically. The invention also aims for scalability, enabling institutions to conduct large-scale examinations with geographically distributed candidates without compromising examination integrity. A further key objective is to enhance the user experience through a streamlined, responsive, and secure system with minimal interruptions during the examination session. The invention also aims to reconcile rigorous monitoring and data privacy by collecting only relevant, encrypted data and adhering to strict data security protocols.Furthermore, the invention is intended to provide institutions with usable analyses and reports derived from candidate behavior during examinations, enabling continuous improvement of examination methods. Finally, the invention aims to create a standardized and tamper-proof online assessment ecosystem that ensures trust and transparency for examiners and candidates and is suitable for academic, professional certification, and corporate audit environments. BRIEF DESCRIPTION OF THE FIGURE

[0017] These and other features, aspects, and advantages of the present invention will be better understood if the following detailed description is read with reference to the accompanying drawing, in which the same symbols consistently represent the same parts. The following applies: Fig. Figure 1 shows a block diagram of ProctorAI: A device-integrated platform for secure and fair online reviews.

[0018] Experts will also recognize that the elements in the drawing are shown for the sake of simplicity and are not necessarily to scale. For example, the flowcharts illustrate the process by highlighting the main steps to enhance understanding of the aspects of this disclosure. Furthermore, with regard to the design of the device, one or more components of the device may be represented in the drawing by conventional symbols, and the drawing may show only the specific details relevant to understanding the embodiments of this disclosure, so as not to clutter the drawing with details that are readily apparent to those skilled in the art after reading this description. Detailed description of the invention

[0019] For a better understanding of the inventive principles, reference is made below to the embodiment shown in the drawing, which is described in specific terminology. However, this does not limit the scope of the invention. Changes and further modifications of the illustrated system, as well as further applications of the inventive principles, are possible, as would normally occur to a person skilled in the art in the field of invention.

[0020] It is clear to the person skilled in the art that the preceding general description and the following detailed description are exemplary and explanatory of the invention and are not intended as a limitation of it.

[0021] References in this specification to “an aspect”, “another aspect”, or similar expressions mean that a particular feature, structure, or property described in connection with the embodiment is included in at least one embodiment of the present disclosure. Therefore, occurrences of the expressions “in one embodiment”, “in another embodiment”, and similar expressions in this specification may all refer to the same embodiment, but need not.

[0022] The terms "includes," "include," or other variations thereof are intended to cover non-exclusive inclusion, such that a process or method that includes a list of steps may not only contain those steps but may also include other steps not expressly listed or inherent in such process or method. Likewise, the statement "includes..." in the case of one or more devices, subsystems, elements, structures, or components does not, without further limitations, preclude the existence of other devices, subsystems, elements, structures, components, or additional devices, subsystems, elements, structures, or components.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as understood by a person skilled in the art in the field of the invention. The system, methods, and examples provided here serve only for illustration and are not to be construed as a limitation.

[0024] Embodiments of the present disclosure are described in detail below with reference to the attached drawing.

[0025] In Fig.Figure 100 is a block diagram of an artificial intelligence-based system for secure and fair online reviews. System 100 comprises: a tamper-proof examination processing device (102) that includes at least one high-resolution camera, a directional microphone array, a secure processor (102a), and a non-volatile storage medium; a facial recognition module (104) running on the secure processor and configured to continuously authenticate a subject by comparing live facial features with pre-stored biometric templates; and an eye-tracking module (106) that detects the angular displacement of the subject's head and eyes relative to the screen and generates deviation signals when thresholds are exceeded.an object detection control module (108) configured to analyze video images captured by the camera to identify unauthorized devices, printed matter, or other persons in the field of view; an audio anomaly detection module (110) that processes input signals from the microphone array and detects unauthorized speech activity, secondary human voices, or irregular background noise patterns; a probability-based misbehavior detection engine (112) that calculates a misbehavior probability score by assigning weighted coefficients to the outputs of the face detection module, the eye-tracking module, the object detection control module, and the audio anomaly detection module;a secure test controller (114) that restricts the device to a test mode in which operating system resources not related to the assessment are disabled, thereby preventing switching between tabs or the execution of unauthorized applications; and an automated scoring engine (116) configured to evaluate both objective and subjective responses of the candidates, with subjective responses being processed using semantic and syntactic natural language processing models to generate a score, and with all results and warnings being transmitted to a remote examiner dashboard via an encrypted communication channel (116a).

[0026] The components of this system are implemented using dedicated hardware units that ensure real-time and verifiable operation. The surveillance processing unit includes a high-resolution camera, a directional microphone array, and a secure processor integrated on a hardware board with embedded non-volatile memory for local data storage. The facial recognition module, eye-tracking module, object recognition control module, and audio anomaly detection module are executed on the aforementioned processor using hardware-accelerated computing units such as digital signal processors (DSPs), field-programmable gate arrays (FPGAs), or dedicated AI inference chips to ensure low-latency processing of the sensor data.The secure processor physically enforces cryptographic isolation and the trusted execution of authentication and recognition algorithms through hardware-based encryption keys stored in a secure enclave. Communication interfaces are embedded circuits that support encrypted data transmission over secure channels, while the automated evaluation engine utilizes on-chip machine learning accelerators for efficient scoring and semantic analysis. The entire system therefore operates with integrated, hardware-based modules designed to ensure secure, reliable, and high-performance evaluation processing without relying on purely software- or cloud-based execution.

[0027] In one embodiment, the secure processor of the investigation processing device (102a) comprises a Trusted Platform Module (TPM) and a hardware-based AI accelerator, wherein the TPM is configured to securely store biometric feature vectors in encrypted form, while the AI ​​accelerator locally performs Convolutional Neural Networks for real-time inference, thereby reducing latency and preventing the transmission of raw biometric data outside the device.

[0028] In one embodiment, the facial recognition module (104) uses a convolutional neural network trained with datasets of both frontal and non-frontal poses, and continuous authentication is performed at regular intervals of 30 to 50 seconds, so that temporal deviations in facial orientation do not affect identity verification and identity fraud attempts are detected by successive mismatched images.

[0029] In one embodiment, the gaze tracking module (106) calculates the gaze deviation by extracting orientation points of the eye region using a Dlib-based regression model, calculating yaw and pitch angles, and comparing these with a predefined allowable threshold of 25 to 30 degrees, wherein deviations exceeding this threshold for more than three consecutive seconds generate a warning signal that is classified as a potential treatment error.

[0030] In one embodiment, the object recognition control module (108) is implemented using a YOLOv5 deep learning architecture trained on a user-defined dataset of handheld devices, printed study materials, and features of the human upper body, and wherein the recognition reliability value is dynamically calibrated based on the ambient lighting detected by the device's integrated light sensor to minimize false alarms in low-light conditions.

[0031] In one embodiment, the audio anomaly detection module (110) performs a short-term Fourier transform (STFT) on captured signals, applies Mel-frequency-cepstrum coefficient (MFCC) extraction, and classifies the processed features using a recurrent neural network trained on human voice datasets. This distinguishes potential speech from background conversation, and anomalies persisting for more than 5 seconds automatically trigger a misbehavior flag.

[0032] In one embodiment, the secure audit controller (114) is configured to disable external memory ports, network interfaces other than the secure encrypted audit channel, and all system-level hotkeys. Attempts to overwrite the environment using kernel-level exploits are detected by an integrity monitor running within the device's secure enclave, thus preventing circumvention of the monitoring system.

[0033] In one embodiment, the automated rating engine (116) comprises a semantic relational feature extraction model for subjective responses, wherein text responses are tokenized, normalized and embedded in vector spaces using transformer-based language models, and wherein the rating is performed by calculating cosine similarity ratings with reference model responses, combined with a syntactic rule-based rating to ensure context-relevant scoring.

[0034] In one embodiment, the encrypted communication channel (116a) between the audit processing device and the auditor dashboard uses Transport Layer Security (TLS) 1.3 with mutual certificate authentication, and all audit logs, misconduct alerts, and candidate responses are hashed with SHA-256 before transmission to ensure data integrity. Audit trails are stored in a distributed ledger for post-audit review.

[0035] The detailed description of the invention relates to a device-integrated online assessment platform that combines hardware-based security mechanisms with artificial intelligence (AI)-supported monitoring and assessment techniques, thereby ensuring fairness, reliability, and resistance to abuse during examinations. The invention functions both as a secure supervision device and as an intelligent assessment system, integrating facial recognition, eye tracking, object recognition, audio anomaly analysis, probabilistic decision-making, and automated assessment in a unified architecture. Each module corresponds to the elements defined in the claims, the technical implementation of which is described in the following sections.

[0036] The facial recognition module is implemented using a convolutional neural network trained on extensive datasets of frontal and non-frontal faces. At the start of the test, the candidate undergoes a registration step in which a reference image is captured and converted into a high-dimensional feature embedding. During the test, successive webcam frames are processed through a preprocessing pipeline that includes normalization, histogram adjustment, and facial feature recognition to account for variations in lighting and occlusion. Embeddings of these frames are compared to the reference embedding using cosine similarity or a Euclidean distance metric. If the similarity falls below a predefined threshold over several consecutive intervals, the system flags a potential identity fraud attempt.Continuous verification is achieved by capturing and authenticating the test subject's image every 30 to 50 seconds. This ensures that the same person remains present throughout the entire session.

[0037] The eye-tracking module uses a regression-based detector to extract landmarks of the eye region from the video stream. Once landmarks such as the pupil center and eyelid contours are identified, the yaw and pitch angles of the head relative to the camera axis are calculated. The system establishes a baseline during the first few minutes of the exam, when the candidate should be fully focused on the screen. Any subsequent deviation of more than 25 to 30 degrees that lasts longer than three seconds is classified as suspicious behavior, as it may indicate that the candidate is consulting unauthorized materials or communicating with others. The threshold is carefully calibrated to avoid penalizing natural micro-movements or temporary distractions. An integrated hysteresis function compensates for noise in the angle measurements.

[0038] The object recognition control module is based on a YOLOv5 deep learning model specifically trained on datasets containing mobile phones, printed documents, laptops, and upper body features of multiple individuals. The model processes sequential video frames captured by the device's camera and outputs bounding boxes with confidence levels for detected objects. To reduce false detections due to lighting fluctuations, the system integrates an ambient light sensor that dynamically adjusts the detection confidence thresholds based on ambient brightness. For example, in low light conditions, the system increases the frame-averaging window to improve robustness, while in good light conditions, it allows for higher frame rates to detect transient errors.As soon as an unauthorized object is detected with a reliability above a calibrated threshold, the event is flagged as a misbehavior and a warning is issued to both the candidate and auditor dashboards.

[0039] Audio anomaly detection is achieved by processing the microphone array's input signals using a series of spectral analysis techniques. Raw audio data is segmented into short frames of 20 to 40 milliseconds and subjected to a short-time Fourier transform (STFT). Mel frequency cepstrum coefficients (MFCCs) are then extracted to capture spectral features. These features are then fed into a recurrent neural network trained on labeled datasets of candidate speech, background noise, and multiple voices. The model classifies each audio segment into categories such as single speaker, multi-speaker, or non-human sound. Unauthorized background conversations or external prompts are detected when multiple voice signatures are identified concurrently with the candidate's speech.Persistent anomalies lasting beyond a threshold of five seconds are flagged as malfunctions to filter out temporary background noise.

[0040] The secure audit controller is ensured by a kernel-level integrity monitor within the device's secure enclave. This monitor prevents the execution of unauthorized processes, disables external memory ports, and ensures that only the audit application can access system resources. Any attempt to circumvent these safeguards, such as kernel-level exploits or privilege escalation attacks, is detected by measuring system call integrity and verifying the digital signatures of running processes. Any detected tampering attempts immediately terminate the audit session and are logged in an immutable audit log.

[0041] For grading and evaluation, the system integrates an automated scoring engine capable of processing both objective and subjective responses. Objective responses are directly compared to a reference database, with scores awarded based on exact matches or tolerance-based rules for numerical inputs. Subjective responses are evaluated using a two-stage natural language processing pipeline. In the first stage, responses are tokenized, normalized, and embedded into a vector space using transformer-based language models such as BERT. In the second stage, semantic-relational feature extraction is performed to assess similarity to model responses, while syntactic scoring rules evaluate grammar, structure, and the presence of keywords.Cosine similarity and context-based scoring metrics are combined to determine a score that reflects both semantic accuracy and linguistic quality. This two-tiered scoring ensures that responses are not unfairly penalized for lexical differences, while still capturing contextual correctness.

[0042] The communication system of the invention ensures the secure transmission of exam data and monitoring logs. The device establishes a Transport Layer Security (TLS) 1.3 channel with mutual certificate authentication to the examiner server. All data packets, including biometric verification logs, error alerts, and response scripts, are hashed using SHA-256 before transmission. Additionally, audit trails are anchored in a distributed ledger, creating tamper-proof records that can be independently verified after the exam. This ensures that neither the candidate nor the examiner can subsequently manipulate the exam data.

[0043] In operation, the invention provides an uninterrupted monitoring environment from candidate login to exam submission. Identity is verified upon access and continuously monitored, behavior is analyzed using multimodal AI techniques, anomalies are integrated into probabilistic decisions, and real-time alerts are generated. The device enforces hardware and software restrictions that prevent circumvention, while the scoring engine enables immediate and unbiased evaluation. By combining secure device infrastructure, advanced detection techniques, probabilistic fusion, and automated evaluation, the invention offers a comprehensive solution to the challenges of online remote examinations.

[0044] The invention also relates to natural language processing and automated evaluation methods for assessing candidate responses without human intervention. In a broader sense, the invention lies at the intersection of artificial intelligence, embedded systems, data security, and distance learning, and offers institutions and organizations a reliable platform for conducting large-scale examinations in academic, professional, and business settings.

[0045] The drawing and the preceding description show examples of embodiments. Those skilled in the art will recognize that one or more of the described elements can be combined to form a single functional element. Alternatively, certain elements can be divided into several functional elements. Elements of one embodiment can be added to another embodiment. For example, the sequence of the processes described here can be changed and is not limited to the manner described here. Furthermore, the actions of a flowchart need not be implemented in the sequence shown; nor does it necessarily have to be performed by all actions. Actions that are not dependent on other actions can also be performed in parallel with the other actions. The scope of the embodiments is in no way limited by these specific examples.Numerous variations are possible, whether explicitly stated in the specification or not, such as differences in structure, dimensions, and material usage. The range of embodiments is at least as broad as specified in the following claims.

[0046] Advantages, further benefits, and problem solutions have been described above with reference to specific embodiments. However, the advantages, benefits, problem solutions, and all components that can lead to an advantage, benefit, or solution occurring or becoming more apparent are not to be construed as critical, necessary, or essential features or components of individual or all claims. REFERENCES 100 A basic system with artificial intelligence for safe and fair online reviews. 102 Examination processing device 102a Secure Processor 104 Face recognition module 106 Eye tracking module 108 Object Recognition Module 110 Module for the Detection of Audio Anomalies 112 Probability-based misconduct detection engine 114 Secure Test Controller 116 Automated evaluation machine 116a Encrypted communication channel

Claims

[1] An artificial intelligence-based system for safe and fair online reviews. The system includes: an investigation processing device comprising at least one high-resolution camera, a directional microphone array, a secure processor and a non-volatile storage medium; a facial recognition module running on the secure processor and configured to continuously authenticate a test subject by comparing live facial features with pre-stored biometric templates; an eye-tracking module that determines the angular displacement of the subject's head and eyes in relation to the display screen and generates deviation signals when thresholds are exceeded; a control module for object recognition that is configured to analyze video images captured by the camera in order to identify unauthorized devices, printed materials, or other people in the field of view; an audio anomaly detection module that processes input signals from the microphone array and detects unauthorized speech activity, secondary human voices, or irregular background noise patterns; a probability-based misbehavior detection engine that calculates a probability score for misbehavior by assigning weighted coefficients to the outputs of the face detection module, the eye tracking module, the object detection control module, and the audio anomaly detection module; and an automated scoring engine configured to evaluate both objective and subjective responses from candidates, with subjective responses being processed using semantic and syntactic natural language processing models to generate a score, and with all results and alerts being transmitted to a remote examiner dashboard via an encrypted communication channel. [2] System according to claim 1, wherein the secure processor of the examination processing device comprises a Trusted Platform Module (TPM) and a hardware-based AI accelerator, wherein the TPM is configured to securely store biometric feature vectors in encrypted form, while the AI ​​accelerator locally performs Convolutional Neural Networks for real-time inference, thereby reducing latency and preventing the transmission of raw biometric data outside the device. [3] System according to claim 1, wherein the facial recognition module uses a convolutional neural network trained with datasets of both frontal and non-frontal poses, and wherein continuous authentication is performed at regular intervals of 30 to 50 seconds, so that temporal deviations in facial orientation do not affect identity verification and identity fraud attempts are detected by successive mismatched images. [4] System according to claim 1, wherein the gaze tracking module calculates the gaze deviation by extracting orientation points of the eye region using a Dlib-based regression model, calculating yaw and pitch angles and comparing them with a predefined allowable threshold of 25 to 30 degrees, wherein deviations exceeding this threshold for more than three consecutive seconds generate a warning signal which is classified as a potential treatment error. [5] System according to claim 1, wherein the object recognition control module is trained using a user-defined data set of handheld devices, printed study materials and features of the human upper body, and wherein the recognition reliability value is dynamically calibrated based on the ambient lighting detected by the device's integrated light sensor to minimize false alarms in low light conditions. [6] System according to claim 1, wherein the audio anomaly detection module performs a short-term Fourier transform (STFT) on the captured signals, Mel frequency-cepstral coefficient extraction (MFCC), and classifies the processed features using a recurrent neural network trained on datasets of human voices. This distinguishes the candidates' speech from background conversations, and anomalies lasting longer than 5 seconds automatically trigger a misbehavior flag.

Images