From ransomware to botnets, malware takes seemingly endless forms, and it’s forever proliferating. Try as we might, the humans who would defend our computers from it are drowning in the onslaught, so they are turning to AI for help. There’s just one problem: machine-learning tools need a lot of data. That’s fine for tasks like […] 
From ransomware to botnets, malware takes seemingly endless forms, and it’s forever proliferating. Try as we might, the humans who would defend our computers from it are drowning in the onslaught, so they are turning to AI for help.
There’s just one problem: machine-learning tools need a lot of data. That’s fine for tasks like computer vision or natural-language processing, where large, open-source data sets are available to teach algorithms what a cat looks like, say, or how words relate to one another. In the world of malware, such a thing hasn’t existed—until now.
This week, the cybersecurity firm Endgame released a large, open-source data set called EMBER (for “Endgame Malware Benchmark for Research”). EMBER is a collection of more than a million representations of benign and malicious Windows-portable executable files, a format where malware often hides. A team at the company also released AI software that can be trained on the data set. The idea is that if AI is to become a potent weapon in the fight against malware, it needs to know what to look for.
Security firms have a sea of potential data to train their algorithms on, but that’s a mixed blessing. The bad actors who make malware are constantly tweaking their code in an effort to stay ahead of detection, so training on malware samples that are out of date could prove an exercise in futility.
“It’s a game of whack-a-mole,” says Charles Nicholas, a computer science professor at the University of Maryland, Baltimore County.
EMBER is meant to help automated cybersecurity programs keep up.
Instead of a collection of actual files, which could infect the computer of any researcher using them, EMBER contains a kind of avatar for each file, a digital representation that gives an algorithm an idea of the characteristics associated with benign or malicious files without exposing it to the genuine article.
This should help those in the cybersecurity community quickly train and test out more algorithms, enabling them to construct better and more adaptable malware-hunting AI.
Of course, making the data set open for anyone to use could also prove a liability if it were to fall into the wrong hands. Malware creators could use the data to design systems that virus-hunting AI won’t recognize, a problem that Hyrum Anderson, Endgame’s technical director of data science, says the company has thought through. Anderson, who worked on EMBER, says that he hopes the benefits of this openness outweigh the risks. Besides, cybercrime is so lucrative that the people behind malware are already well motivated to keep refining their attack tools.
“The hacker will find an example anyway,” says Gerald Friedland, a computer science professor at the University of California, Berkeley.
Read the source article in MIT Technology Review.