AI-powered exam cheating detection system for remote and paper-based exams

Abstract

Claims (s) An intelligent exam proctoring system (100) for identifying cheating behavior in online and paper exams, the intelligent exam proctoring system comprising: • Video and audio recording module (102) for recording real-time video and audio recordings of the examinee through at least one camera and one microphone when the examinee takes an exam; • a preprocessing unit (104) adapted to normalize, denoise and synchronize the video and audio data to form a structured input stream; • a behavior recognition engine (106) operable to process the structured input stream using one or more machine learning models to detect the subject's behavior, including gaze direction, head movement, body posture, object presence, person detection, vocal activity, and the like; • a suspicion assessment device comprising a warning device (108) arranged to assign a suspicion assessment according to detected behavior and to transmit one or more warnings from a group consisting of yellow, orange and red warnings when the suspicion assessment has exceeded a predetermined threshold; • a proctor dashboard user interface (110) for displaying real-time session information, such as examinee status, warning levels, and video data, to a human proctor. • a compliance and security layer (112) adapted to enforce data security, access control and audit logs; wherein The system is designed to run on commercially available hardware and can be used in remote, in-person or hybrid testing environments.

Description

Field of the invention

[0001] The present invention relates to intelligent proctoring systems and, more particularly, to real-time AI-based systems for detecting cheating behavior in digital and paper-based exams using computer vision, audio analysis, and behavioral analysis. Background of the invention

[0002] Academic integrity is the cornerstone of educational systems worldwide and directly impacts the reputation, quality, and validity of institutions' certifications. Tests, whether online, in a classroom, or in a hybrid format, are the most common way to assess students' understanding of material and performance. However, with the rapid proliferation of digital learning tools, fully decentralized distribution models, and asynchronous exams, the difficulty of keeping exams secure and thus preventing cheating has increased exponentially. Traditional proctoring approaches, which rely largely on human supervision, are no longer sufficient to handle the complexity, volume, and contextual diversity of today's exam scenarios.

[0003] Human proctors are typical in traditional in-person exams, where they physically observe test takers to prevent cheating. While this method is somewhat successful, it also has disadvantages such as fatigue, inattention, and poor observation skills, especially in large classrooms or hallways. It is simply impossible for a single proctor to monitor several dozen or hundreds of students simultaneously and with high quality. Furthermore, minor or planned cheating—such as copying, messaging, or using prohibited sources—can easily go unnoticed.

[0004] The rise of online and remote exams due to global events such as COVID-19 has exacerbated the challenges of traditional proctoring methods. To address this need, various digital proctoring systems exist, offering features such as screen recording, keyboard and mouse logging, browser locking, webcam video recording, and biometric analysis. However, these mechanisms often do not account for activities outside of the screen area, such as environmental manipulation or interactions with unauthorized individuals. Furthermore, these tools tend to produce a high false positive rate (innocent behavior mistakenly perceived as suspicious) due to overly strict rules or inadequate context. The resulting lack of confidence and anxiety further exacerbate students' anxiety and distrust.

[0005] Current systems are also heavily hardware-based. Most require multi-camera rigs, a special facial recognition sensor, or a fast, uninterrupted internet connection. These requirements particularly disadvantage students in poor or developing countries, where internet connections and access to modern hardware are unreliable or nonexistent. In some cases, students with strong academic performance may be discriminated against due to the technical requirements of current proctoring systems.

[0006] An even more glaring gap is the lack of tools for paper-based assessments. Handwritten exams are common in certain disciplines such as mathematics, engineering, art, architecture, and linguistics, as well as in institutions and regions where digital access is not realistic. Students do not interact with a screen during such exams, making screen-centric proctoring inapplicable. Furthermore, students frequently glance down at their notes or look askance while reflecting—such behavior, if misinterpreted by digital proctoring algorithms, could be falsely labeled as "suspicious." Current proctoring systems are less advanced and context-aware and cannot be effectively deployed in such offline, real-world environments.

[0007] Pocket Archive: At the same time, student concerns about data privacy and the ethical aspects of data use are growing. Numerous students and advocacy groups are raising objections to the continuous use of video surveillance, opaque decision-making algorithms, and the indefinite storage of personal data. The legal provisions of the GDPR, FERPA, and other national laws impose extremely strict requirements on the collection, processing, storage, and deletion of data. As a result, organizations are increasingly wary of using invasive surveillance solutions without public guidance, consent, and data retention.

[0008] To address these diverse and complex challenges, there is an urgent need for novel, intelligent, curriculum-based, and flexible exam proctoring systems that operate transparently with both digital and paper-based exams. These systems must be high-tech, operationally scalable, respect student privacy, and require few complex or expensive hardware dependencies. Furthermore, they must distinguish between normal / benign student actions and genuinely suspicious actions (i.e., non-false positives) through contextual abstraction (i.e., context awareness) to reduce the number of false positives and ensure a reliable and fair assessment environment.

[0009] Behavior in the room is analyzed in real time using standard webcams and audio feeds. The system uses state-of-the-art computer vision, facial recognition, eye tracking, object recognition, and audio signal processing to identify a wide range of suspicious behavior—from unauthorized entry and use of unauthorized objects, to repeated looking away, speaking despite apparent silence, and even leaving the frame. Crucially, the system requires neither digital testing nor execution on the monitoring device. It is fully scalable and adaptable to the specific testing and audit methodology; whether CB or PB systems are used, the institution wants to maintain the integrity maintenance process within the same process.

[0010] At the core of the invention is a dynamic three-stage warning process—yellow (low suspicion), orange (medium suspicion), and red (high suspicion)—that gradually triggers alarms based on the frequency and severity of the observed behavior. This tiered model is designed to be fair to students, as it gives students an opportunity to correct their behavior after an initial warning and only escalates if the misconduct is repeated or becomes more serious. Warnings can be presented to students visually on screen or audibly and simultaneously displayed on a remote proctor's console, allowing them to see many of the students being proctored in real time.Each alert is recorded with a timestamp as evidence, facilitating transparent reconciliation after the review and ensuring that decisions to determine guilt are made under human oversight.

[0011] Accessibility and Deployment Flexibility One of the main objectives of the present invention is to provide slurry composition with the lowest possible latency. It can be used as a tool in learning management systems, as a standalone application, or in a network test environment. The architecture is designed for efficient, low-latency processing on standard hardware such as laptops, tablets, or external webcams. Furthermore, the system meets the highest data protection standards, and integration into the solution can be customized to meet customer needs. These include options for data storage, local or cloud processing, and consent management.

[0012] The present invention represents a critical advancement in educational technology by providing a scalable, adaptive, and ethical approach to ensuring exam integrity. It addresses the shortcomings of traditional proctoring systems with first-generation AI systems through a comprehensive, interactive, and behaviorally intuitive approach that respects both the interests of the institution and the rights of students. Combining the advantages of automation, human supervision, and simple design, this AI-powered system represents a comprehensive solution that can revolutionize the administration and assurance of academic exam integrity in the digital age and beyond. Summary of the invention

[0013] The present invention proposes a comprehensive, AI-centric proctoring system that can detect, prevent, and report test-takers' cheating activities—regardless of whether the exam is administered remotely via computer or paper in a classroom. Unlike traditional proctoring systems that only consider the user's screen activity or involve additional devices, we utilize an intelligent, behavior-aware approach that leverages standard webcams and audio inputs to enable room-wide monitoring and real-time behavioral analysis. This invention offers a unique combination of computer vision, machine learning, and decision-scoring algorithms to accurately detect the legitimacy of individual test-takers' behavior in a context-aware manner.

[0014] The device essentially records live video and audio of the examinee's environment using standard recording equipment, including a standard video camera and microphone. This data is analyzed in real time to detect anomalies during the examination. In particular, the system detects and interprets a variety of suspicious events, such as leaving the camera frame, frequently looking away from the camera for extended periods (which may indicate suspicious activity), detecting unauthorized electronic devices such as smartphones, smartwatches, calculators, or other tools, the presence of other people in the frame, and verbal communication during silent examinations. The invention utilizes complex computer vision, including facial recognition, head pose estimation, eye tracking, object detection, and person recognition, as well as natural speech and acoustic analysis, to detect voice anomalies.

[0015] The invention also includes a three-level (color-coded) warning mechanism that continuously cycles through possible responses related to the observed phenotypes, their persistence, and severity. The yellow alert indicates a minor / first violation, such as briefly leaving the screen or looking away for a prolonged period. It primarily serves as a mild measure to remind the examinee to change their behavior. The orange alert indicates repeated or more serious violations, such as persistent unacceptable behavior despite previous warnings or a first-time moderate violation, such as another person entering the screen. The red alert, with the highest severity, is issued for serious or repeated violations (e.g., unauthorized use of the device, prolonged absence from the screen, confirmed collaboration with another person).When the red limit is reached, the system can pause the exam, lock the user in the user interface (for digital exam access), immediately notify the proctor, and securely record all relevant video and audio data for post-exam review and assessment.

[0016] To enable continuous operation, the invention includes an intelligent and active proctoring dashboard that allows a human proctor to screen multiple candidates simultaneously. Each candidate is presented with a real-time status indicator based on the current alert status: green (everything's fine), yellow (hey, what's going on here?), orange (uh-oh, we have a problem), or red (hey proctor, candidate needs attention). Upon an alert, the dashboard provides an immediate display of associated incident data—from time-stamped frames and audio clips to AI-based reasoning behind the alert decision. This feature enables teachers and administrators to identify potential violations in context, limit unfounded claims, and promote fair decisions.

[0017] Another key enhancement to the system is its customization, based on a suspicion scoring engine that calculates the cumulative suspicion score for each examinee throughout the exam session. Multiple detectors determine the severity and confidence level of each detected behavior, which are then combined and analyzed to decide when (if at all) to raise an alert and how to communicate it. This model gives the system the flexibility to adapt to different exam types, policies, and organizational preferences. For example, cut scores can be set to account for minor interruptions during a low-risk exam, and more stringent policies can be applied to the same cut score but for a high-risk licensing exam.The scoring engine is also based on a decay logic that can cause a suspicion score to gradually decrease over time if a student maintains consistent compliant behavior, thus promoting fairness and the child's behavioral recovery.

[0018] Crucially, the system is hardware and format-agnostic, allowing it to adapt to diverse exam situations. It can work with online exam platforms or be used as a standalone monitor in classrooms, lecture halls, or remote home exams. For written exams, the webcam captures the candidate's upper body and desk surface—a screen-independent solution. However, for online exams, it can also be used to monitor both on-screen and off-screen activity! Notifications can be sent via on-screen messages, audible alarms, or linked monitoring dashboards, depending on the application.

[0019] The system was designed to protect students' rights. In compliance with global data protection laws, it offers configurable privacy and compliance controls. Organizations can choose between local or cloud video processing, set data retention periods, require consent-based access to review results, and anonymize exam transcripts. The building system follows military-grade methods (e.g., NSA-approved), so no one can eavesdrop or hack your data. All logins are encrypted, and all video and audio clips can be saved, exported, and deleted under the guidance of an administrator. Furthermore, human-in-the-loop review processes are integrated throughout the system, triggering an AI-generated alert for review and appeal by a qualified employee.

[0020] By dynamically combining real-time automation with post-test human review, this invention provides a scalable, context-aware solution that can be integrated into educational systems of any size—from small classrooms to Massive Open Online Courses (MOOCs), certification exams, or national testing centers. It helps ensure academic integrity by preventing cheating through warnings and providing a record of accessed content in case of suspected dishonesty.

[0021] In summary, the invention makes a significant contribution to academic testing and digital authentication by providing a live, AI-based, multi-format, and ethically sound exam proctoring solution. Through the novel use of behavior modeling, adaptive delivery, and intelligent alerting, the limitations of existing systems are overcome and distrust is eliminated. Efficiency and fairness are improved by leveraging the literature on SE testing. Detailed description of the invention

[0022] Fig.shows a block diagram of an AI-based exam cheating detection system. The system of the present invention, hereinafter referred to as system 100, comprises a combination of interconnected hardware and software systems that collectively capture, preprocess, analyze, and intelligently interpret test takers' behavior during online or written exams. System 100 is designed to address a wide range of testing conditions (from home and remote exams to institutional exams) and enhances academic integrity through real-time monitoring, adaptive alerts, and auditable evidence logging.The core elements of system 100 include the video and audio acquisition module (102), the pre-processing unit (104), the behavior recognition engine (106), the suspicious case assessment and warning system (108), the proctor dashboard interface (110), the compliance and data security layer (112), the cloud integration and reporting interface (114), and the hardware optimization layer (116).

[0023] Video and Audio Input Module (102): This module is designed to continuously record multimodal data streams from the examinee in an environment using common consumer hardware (webcams, microphones). The video feed provides facial orientation, gaze, head and body posture, hand position, and interactions with the examinee's workspace. A calibrating field of view includes the desk surface and the immediate surroundings for detection of unauthorized objects or gestures. At the same time, the audio subsystem records background noise, voices, and acoustically unwanted echoes, allowing whispers, distant voices, or irregular sound patterns to be detected. Adaptive frame rates, dynamic gain control, and multi-device synchronization are supported by the module. 102 for use in remote and classroom environments. 1. Preprocessing Unit (104): Raw audiovisual data is input to the preprocessing unit 104, where it is normalized and enhanced to ensure compatibility with downstream AI analysis modules. This processing consists of several steps: using a series of computer vision filters and enhancement models, neural-based noise reduction, brightness adjustment, and spatial motion stabilization. Facial features and hands are located and tracked on each frame. In the case of audio, it filters out background noise, amplifies speech, and analyzes the waveform for presence and continuity. This step ensures that input variations due to different hardware or settings do not affect analytical precision. 2. Behavioral Detection Engine (106): The pre-conditioned inputs are processed by engine 106 in real time and utilizes deep learning models and rule logic to detect behaviors that could indicate cheating. Gaze direction, head posture, and facial orientation were taken into account to exclude long periods of gaze directed away from the examination surface or many slight gaze shifts to the side. The engine monitors body movements and posture changes so that frame exits or unusual body movements can be checked during gameplay. Object detection models are used to fill in mobile phones, notes, or other prohibited materials. Second-party detection programs search for additional people in the environment and mark them. At the same time, acoustic analysis determines if there is any talking, if there are too many or more than one voice, or other noises (e.g., device clicks, etc.).) indicate unauthorized communication. Engine 106 combines the multimodal signals to create a time-synchronized, continuous behavioral profile of the testing session. 3. Suspicion Assessment and Warning System (108): The behavioral events identified by module 106 are evaluated by module 108 using a cumulative suspicion model. The presence of anomalies is weighted according to their severity and a temporal decay parameter. The module calculates a running suspicion level for each subject, which is continuously adjusted based on behavior. The system then automatically activates one of up to three warning levels if the suspicion level exceeds predefined thresholds: Yellow warning: The first warning level concerns a minimal change in previous behavior or the first use of test areas, for example, temporarily looking away from the screen. Orange Alert: A medium alert related to repeated or moderate violations (e.g., partial device exposure, repeated distraction). Red Alert: An alert level indicating that fraudulent behavior has been confirmed or suspected with a “high probability,” for example, through the detection of a mobile device or a second person. Each alarm is time-stamped and logged. It can be indicated to the student either visually or with a soft audible signal and simultaneously communicated to the supervisor. 4. Proctor Dashboard Interface (110), It should be noted that Dashboard 110 is a control and observation center for human proctors and administrators. It provides a real-time panoramic view of all current exam sessions, from examinee videos and alarm status to suspicious result trends and event timeline. The dashboard includes a color-coded alert system that allows proctors to quickly identify high-risk sessions. Each flagged event is available for review in the integrated video player, annotated and flagged for follow-up. The platform also includes real-time proctoring messages, session termination and escalation workflows. It operates either through a browser-based interface or a secure institutional monitoring console with role-based access control and audit logging. 5. Compliance and Security Layer (112), This layer guarantees that all data collection, transmission, storage, and access in the system 100 complies with data privacy and security laws worldwide. All data is encrypted at rest with AES-256 and in transit with TLS. 1.3.exe\x96s secure Login logs (e.g., multi-factor authentication) are implemented to ensure user authentication. The system also logs all user interactions, model outputs, and alarm triggers, which cannot be changed. These logs also help you comply with GDPR, FERPA, and HIPAA. Institutions can set rules about how long data is stored, when it should be anonymized, and when consent should be obtained to comply with their legal and ethical guidelines. 6. Cloud Integration and Reporting Interface (114). System 100 may additionally include module 114 to facilitate access to cloud-based storage, analytics engines, and institutional knowledge systems. This interface enables real-time synchronization with LMS, SIS, and integrity databases. Centralized monitoring on the cloud across multiple locations enables scalable data storage and secure reporting. Enables the export of session metadata, video clips, and SAR summaries in standard formats (PDF, CSV, MP4) for institutional review boards. Federated learning methods can be added to refine the detection models in collaboration with the anonymized data of various institutions. 7. Hardware Optimization Layer (116), The system is equipped with a hardware acceleration layer that supports GPUs, TPUs, or FPGAs to enable efficient execution in high-traffic environments. This layer accelerates video frame processing, deep learning inference, and parallel session verification. It features dynamic voltage and frequency scaling (DVFS) to control thermal margin and power consumption during long examinations. This layer enables the system to scale from individual student devices to institution-wide monitoring deployments while maintaining latency and detection quality.

[0024] Together, components 102-116 form a digital exam proctoring solution that is accurate, scalable, and ethical. System 100 detects cheating in real time using multimodal sensing and AI-assisted decision-making. It is based on a strong human-in-the-loop system and fully complies with data protection regulations. The system is modular and can be deployed in various educational contexts to ensure academic integrity in both traditional and technology-enabled exam formats.

Claims

[1] Claim(s) An intelligent exam proctoring system (100) for identifying cheating behavior in online and paper exams, the intelligent exam proctoring system comprising: • Video and audio recording module (102) for recording real-time video and audio recordings of the examinee through at least one camera and one microphone when the examinee takes an exam; • a preprocessing unit (104) adapted to normalize, denoise and synchronize the video and audio data to form a structured input stream; • a behavior recognition engine (106) operable to process the structured input stream using one or more machine learning models to detect the subject's behavior, including gaze direction, head movement, body posture, object presence, person detection, vocal activity, and the like; • a suspicion assessment device comprising a warning device (108) arranged to assign a suspicion assessment according to detected behavior and to transmit one or more warnings from a group consisting of yellow, orange and red warnings when the suspicion assessment has exceeded a predetermined threshold; • a proctor dashboard user interface (110) for displaying real-time session information, such as examinee status, warning levels, and video data, to a human proctor. • a compliance and security layer (112) adapted to enforce data security, access control and audit logs; wherein The system is designed to run on commercially available hardware and can be used in remote, in-person or hybrid testing environments. [2] The system of claim 1, wherein the behavior detection engine (106) is further configured to detect unauthorized digital objects, including, but not limited to, mobile phones, printed notes, tablets, and electronic wearables, using object detection algorithms. [3] The system of claim 1, wherein the behavior recognition engine (106) determines that one or more other people are in the vicinity of the subject through the use of person recognition algorithms. [4] The system of claim 1, wherein the audio component of the BDE 106 is configured to detect speech, whispering, and crosstalk using voice activity detection (VAD) and real-time speech recognition. [5] The system of claim 1, wherein the suspicion rating and warning system (108) is configured to decay suspicion ratings based on time without current behavioral anomalies. [6] The system of claim 1, wherein the warning system sends a yellow alert for initial or minor suspicious activity, an orange alert for repeated or moderate activity, and a red alert for severe or repeated behavior. [7] The system of claim 1, wherein the dashboard interface (110) further includes features that enable manual override of system-generated alerts, communication with examinees, and annotation of incidents. [8] The system of claim 1, wherein the compliance and data security layer (112) is adapted to enable AES-256 encryption for data at rest, TLS 1.3 for data in transit, and multi-factor authentication for access control. [9] The system of claim 1, further comprising a cloud integration and reporting interface (114) capable of exporting session metadata, video evidence, and alert summaries to institutional databases or academic integrity review systems. [10] The system of claim 1, wherein the system further comprises a hardware optimization layer (116) adapted to perform real-time video and audio processing through the use of hardware accelerators selected from the group consisting of GPU, TPU, and FPGA.

Images