Online exams are the most preferred mode of exams in online learning environment. This mode of exam has been even more prevalent and a necessity in the event of a forced closure of face-to-face teaching such as the recent Covid-19 pandemic. Naturally, conducting online exams poses much greater challenge to preserving academic integrity compared to conducting on-site face-to-face exams. As there is no human proctor for policing the examinee on site, the chances of cheating are high. Various online exam proctoring tools are being used by educational institutes worldwide, which offer different solutions to reduce the chances of cheating. The most common technique followed by these tools is recording of video and audio of the examinee during the whole duration of exam. These videos can be analyzed later by human examiner to detect possible cheating case. However, viewing hours of exam videos for each student can be impractical for a large class and thus detecting cheating would be next to impossible. Although some AI-based tools are being used by some proctoring software to raise flags, they are not always very useful. In this paper we propose a cheating detection technique that analyzes an exam video to extract four types of event data, which are then fed to a pre-trained classification model for detecting cheating activity. We formulate the cheating detection problem as a multivariate time-series classification problem by transforming each video into a multivariate time-series representing the time-varying event data extracted from each frame of the video. We have developed a real dataset of cheating videos and conduct extensive experiments with varying video lengths, different deep learning and traditional machine learning models and feature sets, achieving prediction accuracy as high as 97.7%.