Machine-Learning Project Sifts Through Big Security Data

As an information-security consultant, Alexandre Pinto spent 12 years helping companies set up difficult-to-configure systems to cull security intelligence from logs and security events.

Yet configuring the systems required months of work and even then needed constant maintenance to enable them to detect the latest threats and pinpoint likely malicious traffic. He realized that while companies may want to monitor their networks for threats, they typically have too few security people to work through data from far too many logs -- a problem that will only get worse as companies seek to sift through more operational data to detect threats. Big data could be the downfall of security if companies don't find better ways of dealing with the growing volumes, he says.

"What chance do we have: We can't find the needle in the haystack as it is now, and now the haystack is 100,000 times larger," he says. "We are going to need help."

To help solve the problem, Pinto has worked during the past six months on a machine-learning system that can take logs and identify traffic that originates from suspicious neighborhoods of the Internet. Dubbed MLSec, the project uses supervised learning algorithms to identify networks that are home to malicious actors. Pinto plans to demonstrate the tool at the Black Hat Security Briefings in July.

The independent researcher started with data from the SANS Institute's DShield project, which gathers firewall logs from participating community members. Pinto trained the system on 1.2 million events from 30 million log entries as well as other data submitted by volunteers. When comparing his results to known blacklists, the machine-learning algorithm appeared to be accurate in 92 to 95 percent of cases. Unlike blacklists, however, the systems does not need to be told which networks are malicious; it creates its own representation of the Internet.

Such a system can help information security managers by more accurately flagging traffic coming from questionable areas of the Internet, says Johannes Ullrich, dean of research for the SANS Technology Institute. In addition, the system can give administrators the best guess of the maliciousness of incoming traffic based on incomplete information, he says.

"It really helps to direct the attention of the security administrator," Ullrich says. "The big-data approach filters the data down to a subset, so you know what is worth looking at."

[Rather than watching for communications between infected systems and command-and-control servers, companies can detect stealthy malware when it attempts to spread. See Researcher To Open-Source Tools For Finding Odd Authentication Behavior.]

For companies with overworked staff, the ability to cull the run-of-the-mill data from the interesting -- potentially malicious -- traffic can be a great benefit. In addition, a machine-learning system can be constructed to adapt far faster than a human as the attackers change their tactics, Pinto argues.

"The model will outperform the expert because the model does not forget the data," he says. "It selectively diminishes the weight of what happened before, as time goes by, but it does not forget it."

Pinto plans to make the system available as a service to anyone to upload their firewall logs. In exchange, the people will get a report that summarizes the findings of the system.

In the end, the more people who use the system, the better the results should be, Pinto says.

"In machine learning, of course, the algorithm is important, but the more data that you throw at it, the better," he says. "This is the perfect fit with data security. The more you are attacked, the better your defenses should get."

Have a comment on this story? Please click "Add Your Comment" below. If you'd like to contact Dark Reading's editors directly, send us a message.