Modern Learning Management System uses intuitive graphical user interfaces so students can easily interact with it. The user interface includes icons, buttons, menus, search bars and also provision for online tools through hyperlinked third-party software tools. This greatly speeds up the learning interface for students and faculty alike. Through improved user interfaces, learning management system (LMS) provide easy-to-use tools for the administration, documentation, tracking, reporting, automation, and delivery of educational courses, training programs, materials or learning and development programs.

Currently, the online examination systems supported by open-source LMS of universities can only support a variety of online exam types, e.g., Subjective, Objective, and MCQs. However, the prevention of unfair means in attempting the online exam can be partially guaranteed for MCQ-based exams through measures such as time limits on a question and scrambling of questions and answer choices for candidates, and that too is only applicable when students attempt the exam in specially controlled conditions, e.g., a dedicated exam center only. In other types of non-MCQ-based exams, i.e., numerical-based, subjective, or theoretical, limited techniques have been developed for Massive Open Online Courses (MOOCs), so it is easier for the candidates to use unfair means in online exams on this pattern.

Under the new Outcome Based Education (OBE) systems [1,3], it is mandatory to judge the candidate’s performance on non-MCQ-based exams to measure their course learning outcomes (CLOs) and program learning outcomes (PLOs) [3]. To alleviate this problem of unavailability of video-based online proctoring in LMSs based on the open-source, popular Moodle architecture, usually 3rd party VCSs are used for online proctoring, while the exam paper and completed exam may still continue to be shared through the official LMS channel. As a consequence, (1) The student has to open two software simultaneously to be able to attempt the online exam; (2) As these 3rd party software are not integrated with the LMS, this effectively creates a hurdle in automated solutions for proctoring, a human still has to manually watch the video feeds of all students. Hence, many automated tasks like face recognition and handwriting match, which could easily be performed if these systems had access to LMS data, are not possible.

Many recently developed 3rd party online examination tutoring software are providing additional support for seamless integration with current learning management platforms of universities like Moodle [67], Canvas [68], and Blackboard [69] through open-source customizable plugins. These include ProctorFree [70], HonorLock [71], and Mercer-Mettl [72].

One of the significant issues identified by university academics is that while taking virtual online classes through any video conferencing systems, due to the absence of face-to-face contact, it is often difficult to track which student is present and actively participating in the lecture or merely logged into the VCS. Videos are often switched off due to privacy concerns, especially for female students or when the internet bandwidth is low. A more dire situation may be another proxy student attending the lecture and attempting quizzes/exams on behalf of an enrolled student.

Similarly, during the conduct of the exams, it is difficult to detect any means of unfair cheating by the students. In fact, the use of unfair means is common, especially for subjective exams. Tarigan et al. [73] conducted an exhaustive literature review for different types of academic misconduct by the students attempting online exams. Rabiha et al. [74], in their literature review, drew the conclusion that the most popular online cheating detection techniques used in popular literature were facial motion and head pose, followed by eye movement or gaze detection. Finally, in the third place were the facial movements and gaze detection for multiple targets, as depicted in Fig. 3.

Automated AI-algorithms will be able to establish the identity of the candidate through advanced Face Recognition algorithms. Similarly, Advanced Handwriting checks on answer scripts submitted by students, which is not possible to be done manually by teachers, given a large number of students. This may be managed by implementing CNN algorithms to implement AI-based surveillance of students attending lectures, submitting assignments throughout semesters, and online examination invigilation.

The CNN algorithm architecture is shown in Fig. 4. In this architecture, during the normal conduction of semester classes, student pictures, and their handwriting samples may initially be pulled from the LMS database as a first instance of a labeled dataset sample. More training data of student pictures and handwriting samples can be generated during the normal progress of the semester to allow the CNN to train up to an acceptable level before its implementation can be allowed with low false positives and false negatives. Initially, to validate this training dataset, occasionally, the CNN will take manual help from the professors to limit any errors in the training dataset. In a standard 50-min lecture duration, a considerable amount of labeled datasets for visual student recognition can be generated.

Similarly, ample amounts of handwriting samples can be obtained from student assignments/quizzes being submitted through university LMS systems. In advanced stages, such student data can also be carried forward to subsequent semesters. The above architecture can determine with high accuracy if a student’s identity can be established reliably from the match of this video/image data and handwriting samples. An anomaly will be appropriately flagged for manual investigation by the college administration. Also, valuable statistics can be gathered about the learning experience of students through online education [75] and an evaluation system [76].

Li et al. [77] developed a multi-index cheating detection technique using an LSTM neural network. Similarly, Samir et al. [78] and Essahraui et al. [79] developed a cheating detection system for onsite as well as online examinations by introducing novel computer vision techniques based on CNN-Facial-Landmarks and LSTM models for the classification, which incorporate the detection of students’ head postures and hand movements during the conduction of an online exam. The algorithm can detect cheating for both single-user and multi-user tracking scenarios.

Yulita et al. [80] developed an efficient classifier based on the combination of AdaBoost and SVM for cheating detection using datasets of human activity detection. They found that the proposed approach was more efficient in the detection of cheating compared to other machine-learning techniques. Li et al.[81] developed an online proctoring application by using two cameras working in tandem for 3D gaze tracking. They developed novel gaze motion formulas and also trained a custom-built object detection classifier to detect cheating targets in exams.

Dilini et al. [82] used a One-Class Support Vector Machine (OCSVM) technique for student gaze-based cheating detection. Gupta et al. [83] have developed a smart Bluetooth-based camera application for online exam cheating detection even if the student is taking the exam at his home using the Deep Learning approach. Rudian et al. [84] indicated security vulnerabilities for digital content generation in online lecture delivery and exam creation using H5P, where often the answer to a question can be easily obtained using the source code itself used to develop the exam. They suggested useful measures that can be used by the instructors in setting the question papers in these situations.

Kasliwal [85] used an open-source packet-capturing application, KISMET, for packet sniffing at the student’s computer to see if the student used any of the prohibited URLs during the exam. Bawarith et al. [86] used student biometric authentication via fingerprints, eye tracking, and motion detection techniques via embedded hardware to determine any suspicious activity by the student attempting the online exam.

Atoum et al. [87] developed a multimedia analytics system that continuously tracks six features for the duration of the online exam; gaze detection, text detection, voice detection, phone detection, active window detection, and user verification.

There are a few AI-based software that also help bridge the gap between online lectures and online examinations. One such software is Examus, which monitors students’ behavioral characteristics during online lectures and provides them with proctoring services for better monitoring in online exams [52]. Table 1 provides a summary of research in this domain utilizing AI-based smart algorithms.

Another challenging issue faced by teachers is that students may submit work completed by another student, passing it off as their own. Sometimes, images of handwritten work are not legible. In this solution, students will attempt the exam using Digital Writing pads, which will make the job of handwriting recognition simpler and faster. Also, such digital stationery will credibly establish the real-time geographic location of the exam candidate throughout the duration of the exam, even in the absence of video surveillance. Liwicki et al. [93] made the case for handwriting detection through pen-enabled displays/tablets for instant handwriting recognition.

AI-algorithms can then be used for real-time detection of handwriting. Similarly, Kumar et al. [88] made a case for using an AI-algorithm for handwritten character recognition for writer identification using Recurrent Neural Networks (RNN) and LSTM networks. KreB et al. [89] proposed a Deep Neural Network (DNN) based technique in which sensor data is first stored in a compressed format in the electronic pen through quantization of weights of the DNN layers. Then, a Connectionist Temporal Classification (CTC) algorithm is used to minimize the link bandwidth. This allows transmitting only a small matrix from the pen to the mobile device on which the decoding is performed. Additionally, to further improve the accuracy of handwriting recognition, the authors added a Language Model (LM) to the decoding algorithm executed on the mobile device. This technique distributes the load on the hardware while making edge computing possible for handwriting recognition.

Respondus [54], a commercially available software that uses Respondus LockDown Browser, keeps open only one browser and disables access to the rest of the endpoint (user computer) environment. Respondus can monitor students’ behavior by using a webcam and perform video analytics to detect behavioral events that can signal cheating during online exams. Similarly, BeyondTrust Privilege Management [92], another commercially available software, is an endpoint security software that combines privilege management, application control, and sandboxing. It helps prevent unnecessary granting of admin rights to users and compromising security by enabling users to log on to the desired applications using their standard user accounts. This is done by limiting traffic to only applications using specified TCP port.

Another but limited solution for securing access control is a virtual private network (VPN). Some already available applications like Sonicwall [94] enable restricted and secure VPN access and firewall solutions for higher education examination purposes.

There is an increasing concern about the shift of control from academic institutions to commercial providers, as investigated by Selwyn et al. [95]. The AI-algorithms can frequently make mistakes or out-of-context decisions as they are only as good as they are ‘trained.’ Henry et al. [96] brought to light a case of a student who was disqualified from an online exam by an online proctoring algorithm, only to be reinstated swiftly when she lodged a formal appeal against the decision, with the professor apologizing for the mistake made by the software. There is a dire need to establish the amount of autonomy academic institutions are willing to give to automated algorithms for all academic misconduct during online examinations while exonerating their own faculty members completely if any student is unjustly marginalized. Also, there is a need for special grievance policies should a student decide to contest a decision made by an automated AI-algorithm.

Computer adaptive testing was established three decades ago by international testing services like ETS for GRE exams. The questions appear from different difficulty level pools and try to zero in on the true academic potential of the student. Newer initiatives [97] follow a two-pronged approach for harnessing AI in effective learning of students using both adaptive and adaptable learning. Another issue that can be solved using this approach is that conventional LMS systems based on Moodle architecture only run the exam timer on the main LMS server. Sometimes, internet disruption during time-limited exams proves disastrous for candidate’s performance in the exam when they miss valuable examination time in the event of client-side computer problems such as internet or power outages. It is necessary to have a seamless examination system in which the adapted examination system can run locally on the client side and an encrypted version of the past time-stamped student responses be saved on the local computer with a local timer operating on the local (client side) computer with candidate’s browser and audio/video logs which could be submitted as soon as Internet service is restored.

Female teachers and students are reluctant to turn on their videos during the conduction of online classes and exams due to privacy reasons that any videos recorded in the class or during exam invigilation could be used by cybercriminals. Due to this, many female teachers opt for an asynchronous mode of online classes, or the conduction of live classes with the video turned off. This creates learning issues for many students. Similarly, female students are reluctant to have their video recorded while conducting online exams. Thus, privacy issues are at the forefront for both female teachers and students, as highlighted in recent research [98]. Student has frequently expressed concerns about online proctoring systems recording private and sensitive information from their computers [99,100]. Online education systems should safeguard these concerns so that female teachers and students are able to share their videos with their privacy concerns addressed using encrypted video streams [101–103] and advanced AI-based algorithms to still be able to carry out their face detection accurately even with the presence of hijab/niqab [104–106]. There is also an increasing concern about ethics when handling students’ and teachers’ private information; this has been discussed in detail by Henry et al. [96].

The other important issue is that the students connecting to such online video conferencing systems connect through different networks with different connection speeds and bandwidths. Turning on high-resolution videos for multicasting puts needless strain on the network, and often, universities advise teachers and students to keep their videos turned off. This again exacerbates the issue of students using unfair means during regular classes and exams. Top automated proctoring solutions commercially require high bandwidth internet connections for satisfactory performance. Pearson VUE online proctoring software requires at least a 3 Mbps connection [107]. Elastic applications with rate adaptation according to network conditions have been developed using efficient transport layer algorithms, as shown in Fig. 5.

Notable works in rate-adaptive video streaming applications include Dynamic Adaptive Streaming over HTTP (DASH) [108], and HTTP Live Streaming protocol (HLS) [109], which adapt the bit-rate based on network conditions. DASH divides videos into segments and applies H.264 [102] or HEVC advanced video coding [110] to encode them at varying quality levels (bit-rates). These encoded segments are then stored on the server. The client selects the most suitable segment to transfer the video based on the current network capacity estimation. Another viable option for real-time video streaming is by using Software Defined Networks (SDN). Service providers can interact with apps and devices through a programmable interface provided through SDN. This gives consumers the ability to negotiate bitrate streams for real-time services and make quick requests for resource allocation for various services. The SDN design offers a method for separating network control and routing duties, leading to programmable and abstracted control that is unrelated to network services and applications. SDN makes it feasible to design flexible topologies and new protocol extensions for networks. Only a few emulators, including Mininet [111] and OpenVswitch [112] are readily available right now.

Like rate-adaptive data transmission schemes to salvage real-time data transmission in the face of internet outages, legacy data compression may also be used, albeit for non-real-time or near real-time applications. KreB et al. [89] proposed a handwriting recognition technique using an electronic pen in a deep neural network (DNN) technique in which sensor data is stored in a compressed format. This allows transmitting only a small matrix from the pen to the mobile device on which the decoding is performed. Similarly, Rajesh et al. [90], in their work, HWRCNet, developed a data-compression-based handwritten characters recognition technique combining CNN and Bi-Directional LSTM (BiLSTM) based RNN.

It is also important to implement AI-algorithms at the edge through the IoT concept [113], which will make the system secure [114], and scalable and reduce the burden on a centralized AI inference engine for real-time cheating detection. Online proctoring is heavily dependent on real-time computer vision algorithms. These advanced computer vision algorithms run on the server-end. In the event of session disconnection due to a real or doctored internet/power outage, the computer vision processing will cease giving an opportunity to students for academic misconduct. If these algorithms are running at the edge, i.e., client-end, then automated proctoring can continue unabated and also notify the central server if any anomaly is detected at any point in time.

Varnavsky [115] investigated the efficiency of Flipped Classroom technology using chatbot-based student interaction. Pareira et al. [91] devised an intelligent system where an AI-based recommender system proposes tasks for student assignments in computer programming courses to avoid plagiarism and truly reflect students’ efforts by monitoring the IDE used to develop the assignment. Initiatives like [97] follow a two-pronged approach to harnessing AI in the effective learning of students. AI-based adaptable and adaptive learning is possible through both a mobile and web interface. A learner can learn what he is capable of learning (adaptive) and in the manner he wishes (adaptable learning). The system, through deep data analytics, determines the contents of the course and the manner in which it should be presented to enhance the learning of the student as per previous attempts in completing assignments and tests. An intuitive UIUX design is also a must to support enhanced user experience.

Future education delivery including that using LMS, will be highly affected by generative-AI based content generation and knowledge discovery tools. The popularly available ChatGPT AI-chatbot to everyone including students is trained on a vast amount of data available from the Internet, and its GPT technology allows it to produce human-like responses to various input prompts. ChatGPT is an advanced language model that utilizes the “transformer architecture” to perform various natural language processing tasks such as language understanding and generation [116]. This unique architecture enables AI-powered chatbots to interpret the relationship between words in a sentence, thereby preserving context and producing responses that are both logical and coherent [117]. In engineering, sciences, and related research disciplines, this leads to technical content generation like experimental protocols, designs, and results [118]. When it comes to research content writing, the ChatGPT-based approach offers a significant advantage in terms of efficiency. Writing papers or a thesis can be a time-consuming task, but with this method, users can save time and effort by receiving instantaneous relevant responses from ChatGPT based on specific input prompts related to their research areas. Technical knowledge and area expertise are crucial for engineering and scientific research paper writing, which can make it quite challenging. However, the use of ChatGPT AI-chatbot can be of great help, as it can generate relevant content based on carefully drafted prompts provided by the researcher, thus saving time and effort in the writing process [119]. Another advantage of utilizing ChatGPT-based AI-chatbot for scientific and engineering research paper writing is the potential to enhance the consistency and precision of the way the papers are written. Publishers like IEEE, Springer, Elsevier, Wiley, etc., who publish technical research papers generally demand a high level of accuracy and consistency from the authors, which can be difficult for human writers (especially those who do not have English as their first language) to maintain consistently. ChatGPT-based techniques can produce a dependable and precise output, decreasing the likelihood of errors and disparities [120].

However, ChatGPT also has some limitations. One of the major drawbacks of the ChatGPT model is its lack of deep understanding and potential to produce erroneous or faulty information. ChatGPT is trained on a vast amount of data (from the internet) that could contain inaccuracies or biases, resulting in the model generating misleading information. Furthermore, the model may not possess the required proficiency to create highly precise technical content, particularly for specialized fields where area-specific advanced domain knowledge is required [121]. As an AI-powered assistant, ChatGPT has limitations when it comes to assessing the credibility of its training data, which can affect the accuracy of the information it generates [16]. Moreover, ChatGPT does not have access to the Internet and currently has limited knowledge of world events after 2021 [122,123]. Therefore, some of its responses may be outdated or inaccurate since knowledge is continuously evolving, and its training data set is outdated. For example, when ChatGPT is required to produce up-to-date references for any specific area, ChatGPT may provide fabricated references that seem plausible but are not backed by real-world sources [122].

One drawback of utilizing ChatGPT in the scientific and engineering disciplines for research paper writing will be its potential to lack the creativity and tone that human writers possess. Writing research papers requires more than just providing accurate and consistent data; it also involves effectively conveying complex information in a clear and simplistic manner. The use of ChatGPT may result in less engaging and effective output due to its possible lack of creative and nuanced writing [122]. Another disadvantage is that ChatGPT-based scientific and engineering written content could be biased as per the training dataset and algorithm design, etc., which is one of the biggest problems in AI-powered chatbots because they lack the basic human touch in writing and hence avoid the biases. As per the principle of ‘garbage-in-garbage-out,’ AI-powered chatbots have the potential to reinforce biases by utilizing large sets of data that already contain these biases [124].

In academia and research, the use of ChatGPT as an AI-powered chatbot with the ability to generate text that resembles human language has sparked several ethical concerns related to academic dishonesty [113,119,125]. Although ChatGPT’s responses are not exact copies of any specific text, they are generated by synthesizing training data, which may lead to producing responses similar to those of existing sources. A recent test [116] has shown that ChatGPT wrote a 500-word essay that had a 45% similarity to existing sources. The danger is that students might use ChatGPT without realizing the potential for plagiarism [121]. Moreover, ChatGPT’s ability to generate research studies and scientific abstracts with fabricated data poses a significant risk to academic integrity [121]. This capability may encourage higher education students to rely solely on ChatGPT when writing academic essays, which leads to academic dishonesty. Furthermore, ChatGPT’s tendency to generate incorrect and nonsensical answers raises the risk of misinformation in scientific publications, making this ethical issue even more serious.

As generative-AI based content generation tools are beginning to emerge, proponents will advocate their use as long it is responsibly acknowledged by the student. The AI-chatbot, ChatGPT, is extremely popular, but academics and researchers all over the world are worried that excessive use of ChatGPT in technical research papers and thesis writing [122] will result in sub-optimal research efforts. Academicians and researchers predict that the uncontrolled use of ChatGPT can be quite catastrophic for the research community in general and the scientific and engineering research fraternity in particular [123]. Some alarming facts show that ChatGPT has demonstrated its capacity to provide satisfactory responses to exam questions in medical and legal fields [118,123]. An empirical study revealed that ChatGPT’s answers to life support exams at a university were, on average, relevant and accurate and more closely aligned with guidelines than previous studies that utilized other AI tools. Given this level of performance, ChatGPT poses a significant threat to academic integrity, particularly in higher education [122].

To overcome the problem of plagiarism, counter AI-detection tools like GPTZero [15] are already in use. These tools can predict if a certain text was potentially generated by generative-AI software by analyzing the linguistic constructs in the text [125]. Moreover, quite recently, Turnitin similarity index checking software has also introduced a new option in their software which checks the AI involvement in the written text [11].

Authenticating the identity and behavioral patterns of the student studying or taking exams in distant (online mode) is one of the biggest future areas of research for LMS development. Advanced image/video processing techniques backed by AI will be required for biometrics-based student authentication and behavior detection [126–130]. Previously available techniques, such as Face recognition, have often been shown to be easily deceived if a photo of an individual is presented to a camera. The new research direction will focus on diverse student biometrics-based user identification such as fingerprint, handprint, and iris detection using commodity webcam or smartphone cameras. Biometrics-based student authentication will be a big step toward the prevention of academic misconduct in online lecture delivery and exams.

User Datagram Protocol (UDP) and Transmission Control Protocol (TCP) are the most commonly used transport layer protocols for elastic data delivery on the Internet. This covers data delivery with different QoS requirements, such as latency-sensitive or non-real-time data delivery. Of these, UDP is preferred for real-time data applications, and TCP is used for stored video streaming. TCP ensures reliable data transmission through variable data-rate on the basis of receiver feedback; it does so by retransmitting any lost data packets, making it unsuitable for any real-time video transmission. UDP uses fixed-rate data delivery irrespective of network conditions. Transmitting lower-resolution video quality or a video with many lost frames may compromise the effectiveness of AI-algorithms used for online proctoring and other tasks. AI-algorithms used for advanced video analytics need to work with higher-quality videos. Modern LMS video conferencing applications will use a combination of TCP and UDP for video transmission. TCP will be used for the control plane for probing the network, and UDP will be used in the data plane for flexible rate data delivery, ensuring acceptable QoS. Newer transport protocols designed for video content delivery, such as those used for online lectures through LMS, are becoming increasingly popular. These protocols are equipped with advanced network probing algorithms, such as PathRake TCP [131], which allow for efficient estimation of available bandwidth, even in the presence of firewalls on the network path. The use of these protocols will greatly improve the user’s quality of experience.

Legacy routing service on the Internet over TCP/IP protocol is only providing the best-effort service for data delivery on the Internet. Currently, LMS applications use these conventional routing services for video conferencing applications for online classes/exam conduction. The result may be sub-optimal as end users may be connected with variable connection speeds. In addition, network conditions may fluctuate with time. Specialized video conferencing applications, such as Zoom, use peer-to-peer data delivery methods to optimize video transmission, which provides additional network probing service to gauge the network conditions to quickly reroute packets in events of congestion on default routes. LMS applications of the future will instead rely on the Software-Defined Networking (SDN) framework, which gives the network the ability to re-route data packets quickly in the event of Network Failure or if the path does not meet the user QoS requirements, providing seamless service offering dedicated QoS without any involvement of LMS application for network probing service. In the future, segregated data services in the LMS applications requiring different QoS may use different routing services in the network. Real-time video delivery in future LMS will integrate smart cameras instead of regular webcams, which could interact seamlessly with SDN controllers [132] directly while the remaining latency-insensitive non-video data service could use conventional routing services. The SDN controllers will work on adaptive routing protocols to optimize QoS [133] for efficient video delivery to be processed by AI-based image/video processing algorithms at the server end.