How AI can spot cheating without breaching student privacy

The COVID-19 pandemic has shut down schools and universities around the world, and institutions have had to swiftly adapt to online lessons and exams. Digital proctoring companies used this opportunity to increase sales of their surveillance tools. According to research by Woldeab and Brothen, students with high test anxiety perform worse in a proctored online setting. But how can we trust students not to cheat during an online exam? 

We know people cheat—it’s human nature. But as the digital transformation of the education sector is catalyzed by the pandemic, students and universities are facing a dilemma: protecting privacy while preventing cheating. Is it possible for universities to conduct online tests without breaching the test takers’ rights? We students at IE University are looking for the solution. 

The problem: the need to spot cheating while respecting privacy

IE University professor Manoel Gadi, an expert in risk and fraud analytics, wanted to conduct his second online Data Science Challenge. He couldn’t use SMOWL, a popular proctor software, due to its limitation of offering long-form, written answers and programming the responses his students required. It was then that Professor Gadi decided to challenge us to create our own solutions to prevent online cheating. Alongside three of my classmates—Reem Hageali, Valeria Zaldivar, and Amaya Hijazi—I created an AI model that uses image classification to spot cheating using a screenshot of the exam candidate’s computer activity.

The solution: AI image classification 

We created the AI-driven image classification system because it is far less intrusive than SMOWL’s face- and voice-detection system. The tool was trained on a sample data set of 590 images to train the model to accurately classify cheating or not cheating. Training data consisted of 400 samples of cheating and 190 samples of not cheating. Images that showed cheating contained screenshots of Facebook, WhatsApp web, Google, and Gmail, whereas images that were considered not cheating contained screenshots of the IE campus, Kahoot, and a programming interface called Jupyter Notebook. 

When put to the test in the classroom, our model had a precision rate of 98% accuracy. The graph below explains the privacy and prevention balance we’re trying to achieve:

Figure 1: Tradeoff between protecting privacy and our AI model’s ability to spot cheating 

Ethical solutions for the future of education 

Our solution was used during the Introduction to Social & Business Analytics final exam. Our model discouraged students from visiting social media sites during the exam and some students tricked the system by using their mobile phones to communicate. Nonetheless, the model’s precision was high and it was overall more effective and respectful of privacy. 

With the rise of online education, there is no perfect solution to spot cheating online, and using AI is just one way to do this. Creating this tool has shown us that whatever solution we implement, it must not cause unintentional harm to students by breaching their right to privacy, while also being efficient enough to not be outsmarted by students, because this could result in e-proctor companies creating more authoritarian tools to monitor students.

The key takeaway from this project is that perhaps the best way to prevent surveillance technology from being widely adopted in online education is to take a behavioral-based approach. Another solution is for universities to reassess how they test students’ knowledge. We’re eager to see what the future holds and look forward to using data and AI to create ethical solutions. 

 

Tharun Komari was born in India but lived in the United Arab Emirates for most of his life. He’s a third-year student in the Bachelor in Data & Business Analytics. He’s passionate about applying artificial intelligence to the fields of finance and football and also loves thinking about the future of humanity. You can reach him on Instagram (@tharunkomari) or LinkedIn.