Security Experts:

DARPA Hopes Machine Learning Technologies Can Improve Security

It sounds like science fiction, but advances in technology has made it possible for computers to look at large amounts of data and make decisions based on available data.

Having computers capable of learning and thinking will have big implications in security, and Defense Advanced Research Projects Agency (DARPA) is investigating ways to make it easier to build that technology.

DARPA launched the Probabilistic Programming for Advanced Machine Learning (PPAML) program on Tuesday to combine new programming techniques with machine learning technologies. While there is a lot of demand for machine learning capabilities, every new application requires a significant amount of effort to design and develop because there aren't many tools or systems available to make the process a little bit less challenging. This is where DARPA's PPAML program comes in.

Machine learning refers to the ability of computers to understand data, analyze and manage results, and infer insights from disparate sources of data. A lot of research and development is focusing on giving computers the capability to look at collected data, analyze the information, and draw relevant insights in order to develop new skills, Peter Van Der Made, current chief scientist at Australia's vWISP and former chief scientist at IBM, said while promoting his book on artificial intelligence.

"By producing computer chips that allow computers to learn for themselves, we have unlocked the next generation of computers and artificial intelligence," Van Der Made said.

DARPA's PPAML program seeks to increase the number of people who can successfully build machine learning applications and help the experts become more effective. Future machine learning projects shouldn't require people to know everything about machine learning to build useful applications, Kathleen Fisher, a program manager at DARPA, said in a statement.

The program will help create more economical, robust and powerful applications that need less data in order to generate more accurate results. The program will hopefully reduce the current barriers to machine learning and encourage innovation, productivity, and effectiveness. A successful solution will involve contributions from many areas, including statistics and probabilities modeling, approximation algorithms, machine learning, programming languages, program analysis, compilers, high-performance software, and parallel and distributed computing.

The 46-month long program will run in three phases between 2013 and 2017, according to DARPA.

"We want to do for machine learning what the advent of high-level program languages 50 years ago did for the software development community as a whole," said Fisher. The DARPA Special Notice document describing the specific capabilities sought is available on its site. 

From a security perspective, there are many potential applications of machine learning, and some are already available in the market in some limited forms. Email spam filtering technology is one such example. At the moment, though, security technologies are currently able to do some data analysis, but it cannot make the final decisions. Security technologies can take on some of the tasks and free up the human experts to focus on looking at the data and drawing conclusions.

DARPA's latest program aims to work on technology and automated tools that will make it possible--and easier--to teach a computer than to program it.

Related Reading: Making Systems More Independent from the Human Factor  

Subscribe to the SecurityWeek Email Briefing
view counter
Fahmida Y. Rashid is a Senior Contributing Writer for SecurityWeek. She has experience writing and reviewing security, core Internet infrastructure, open source, networking, and storage. Before setting out her journalism shingle, she spent nine years as a help-desk technician, software and Web application developer, network administrator, and technology consultant.