Machines vs Malice: Artificial Intelligence Takes On Hacking

To catch a thief, set a thief? JL

Quentin Hardy reports in the New York Times:

Sometimes the best way to stop a bad machine is with a lot of good machines.

Sometimes the best way to stop a bad machine is with a lot of good machines.

Several companies are applying the techniques of artificial intelligence, or A.I., to the world of security, and they are using a whole bunch of machines strung together in so-called cloud computing networks to do it. Originally the province of university researchers and now one of the ways Google and other companies figure out what is going on across the web, A.I. technology is being employed by security companies who say they can beat criminals by using many of the same strategies.

Much as Google examines websites for significant information and watches the behavior of people searching and surfing the web, A.I. security companies look for malicious sites or try to examine and predict the behavior of malware, which is software meant to cause problems.

“We’re looking at about 200,000 samples of malicious code a day, so we can guard maybe 11 million events in a microsecond,” said Tomer Weingarten, the chief executive of a computer security company called SentinelOne. To stay on top of that volume, Mr. Weingarten said, requires the equivalent of 10,000 computers.

As computing becomes more pervasive, traditional defenses are proving inadequate. For example, the firewall, which was once an effective safeguard on the perimeter between a corporate network and the world, is now problematic: It has become harder to say where systems begin and end as they become connected to more and more things. In 2013, Target was hacked when criminals entered the main servers through software for a company heating system that was managed by a contractor.

More recently, “sandboxes” have been developed that temporarily isolate incoming programs and files to see if they try something malicious. In response, hackers have written code enabling malware to recognize that it is being quarantined — sometimes by contacting a computer’s operating system directly — so it does not take any suspicious action until it detects that it has been released.

Every day, SentinelOne’s computers scour the many listings worldwide of known malware and attack codes, which are publicly posted by government agencies and private security organizations. Using machine learning, an A.I. technique of pattern mapping, the computers then look for similarities with known techniques and try to identify similar behaviors that precede attacks.

That information is then loaded into computing “agents” that are inside its clients’ computers. The agents observe events inside a computer almost the moment they occur. If, for instance, a so-called ransomware program starts to encrypt a user’s files (to lock up the computer, which will be freed only once the owner pays a ransom), the agent will isolate the program and notify the system administrator.

Often, it can also undo whatever damage was caused by reverting the few files that were affected to an earlier state.

“Sometimes it’s easy to see malicious behavior — no legitimate application would just start encrypting everything,” Mr. Weingarten said. “Other times, they are ‘spraying the heap,’ looking for all the commands being queued up in the computer so they can rewrite the system and insert their code. Normal applications don’t do these things.”

Every piece of malware also has its own biography within the system. Mr. Weingarten recently called up a program called Troldesh, which was first observed on the evening of April 9. It created files on the infected computer, then changed the files and notified a server in Russia that it was ready.

“This starts to look suspicious,” Mr. Weingarten said. Signals can be bounced around, so it is hard to say just where Troldesh originated. It also communicated with machines in Hungary, Austria and Germany.

Troldesh was identified and stopped, but a hacker could reuse much of the code in other malware. That is why A.I. tries to learn hackers’ rules and habits.

Another challenge in protecting today’s computer networks is how poorly understood much of the world’s software is. “There are 600 million individual files known to be good, and a malware universe of about 400 million files,” said Lawrence Pingree, an analyst with Gartner. “But there’s also 100 million pieces of potentially unwanted adware, and 200 million software packages that just aren’t known. It takes a lot of talent to figure out what’s normal and what isn’t.”

The process, which he calls “endpoint detection,” looks at and acts on what goes on in individual machines.

Many of the same techniques can also be used on other kinds of bad online behavior. Carlos Guestrin, a well-regarded expert in machine learning, is chief executive and co-founder of a company called Dato. In addition to traditional A.I. businesses like figuring out shopping preferences, he started looking at fraudulent behaviors.

“We caught spam with machine learning by looking at sequences of words, now we look for the code in a virus, like DNA, that makes it do unusual things,” Mr. Guestrin said. “With human fraud, you look for relationships about who sends money to who, or who is hiding fraudulent transactions. If a finite number of people keep sending each other money, they’re probably trying to look like legitimate businesses.”

G2 Web Services, based in Bellevue, Wash., helps banks figure out if a website is fraudulent or is selling contraband. Using Mr. Guestrin’s product, coupled with human experience, on hundreds of millions of sites, G2 improved its ability to predict fraud and crime by 13 percent. Over millions of transactions, that amounts to quite a lot.

G2 can also flag prohibited content, like child pornography, which exists on about 1.5 percent of all merchant websites. Sometimes a criminal will put a link to, say, a store for illegal growth hormones in an otherwise honest site, without the merchant’s ever knowing about the link placement. Another use for A.I. is spotting “transaction laundering,” in which an illegal business tries to appear legitimate by processing transactions through a legal site.

The company is making strides against cybercrime, since “the guys who run these illicit sites are also into viruses and malware,” said Alan Krumholz, principal data scientist at G2. “It’s a cat-and-mouse game. They go from one business into another.”