I read in today’s Sydney Morning Herald the continuing story of universities in Australia fighting a constant battle with cheats in exams and assessments. Today’s story reveals that there’s not just a problem with plagiarism in essays, but also students paying impersonators to sit their exams for them:
 | ||
University students are increasingly paying impersonators to sit their exams or smuggling in technology to help them cheat, while other students cannot be trusted to sit in sloping auditoriums because of their willingness to copy answers in multiple choice tests, a new report reveals. | ||
 | ||
The story focuses on one investigation report from an Australian university, but the problem is likely to be happening across a broad range of universities, in Australia and internationally.
And I realised that two weeks ago I wrote about a way to tackle this, in a way that’s cost effective, quick to implement and simple to do. The answer lives inside my story about “Making machine learning in education easier for every day users” - and something I’d been talking about with customers recently…
CheatorNot.net
The simple summary of “Making machine learning in education easier for every day users” is that we’ve developed a series of recipes to help build intelligent services, called Project Oxford, and one of the recipes is to do face recognition - and you can quickly build it into your own app, website or software. The services take the complexity of machine learning tasks, hide all the detail, and let you just perform a simple task - in this case “Are these two pictures of the same person?”. And it is very simple for a developer to use it, because it’s based on our Microsoft Azure Machine Learning service.
Mat Velloso, one of our developers, built it into a website called TwinsOrNot.net, which lets anybody do the comparison by uploading two photos of your own, or finding two photos with Bing Image Search.
What is amazing is that Mat built this sitting in a hotel room in the Czech Republic, in one evening, thousands of miles from home, in just four hours. You can read his story of how he created twinsornot.net here, and how it went from a geeky evening-to-kill hotel-room project to a massive viral success (it went from zero users to 75,000 within 7 hours of being demonstrated at a local conference, and a million hits within days).
The user experience is really simple - you pick your two photographs, and it gives you a percentage probability that the two are the same person.
The image below is as close as I could get within our own local Education team - apparently there’s a 66% resemblance between Keith and Jason (which really isn’t that close).
So although it would probably take a human a bit longer to do the checking, the software can instantly tell us that we need to check out Jason if he turns up in the exam hall pretending to be Keith!
The system also makes allowance for every day differences - different lighting conditions, different styles of photo - even different facial hair. So the two photographs of me below were taken 2 years apart, one with full beard (yeah, I know, I never did that again!) and one without. And yet it knew that both photographs were of me.
Although TwinsOrNot.net was built as a fun website, exactly the same services could be used to build an app that runs on an exam moderator’s phone, or on a laptop at the entrance to the exam hall, that compares a student’s ID card photo to the person entering the exam, and in real time reports to the proctor whether they are the same person, or there needs to be more checking done.
If Matt could build this in an evening, then could the same be possible for a university? Well, they already have the components - almost all use the Microsoft Azure cloud services already; they’ve got laptops with webcams and they have got student ID photos. So all it needs is for a developer to spend a few hours building a prototype, and then they could try it in an exam hall by the end of the week. And, just as importantly, they could be ahead of the newspaper headlines within hours…
This isn't a perfect solution that could completely solve the problem, but (a) this could be done quickly at low cost and take a step forward against cheating and (b) doing it will improve the detection of cheating without adding a huge workload for staff. It’s not designed to give 100% assurance, but out of 100 students it would provide a way to highlight the 5 people that need a bit more checking by a moderator.
I think that the battle to combat exam cheats is similar to the battle against computer viruses - it’s a game of cat and mouse, and it’s a constant game of improvement iteration at high speed, and this suggestion is another step forward in the game…
Do you want to build a CheatOrNot website or app?
1) Well, the website domain is available, because I only just thought of the name…
2) Listen to Matt Velloso on the MS Dev Show sharing his experiences
3) Read Mat’s blog, where shares the links he used for building TwinsOrNot.net
4) Mat's even shared his source code to get you started!
Once you’ve built a prototype, let me know how it goes!