Deep down, if I think about who I am, I’m a scientist who loves to solve problems. If you think about cybersecurity, its problems are unique in that we are not only competing against industry competitors, we are also competing against the adversaries behind the cyber-attacks. My recent keynote at MPOWER17 Las Vegas focused on the problem of out-innovating these adversaries.
A year ago, I introduced a framework illustrating how defensive technologies are effective over time based on the innovation competition between defender and adversary. It shows that a defensive technology works best when it is first deployed. At that time, the threat it is designed to address is well-understood. Over time, however, defenders are incentivized to develop more and more countermeasures that will eventually degrade the technology’s efficacy. We have seen this play out with spam filters, sand boxes and numerous other defensive measures.
At McAfee, we have thought a lot about how we can use this cycle of attacker-defender innovation to benefit customers.
First, we take a platform approach by making it easier for you to install and maximize the value of the technologies within your environment. Value could mean things such as technology teaming enabled with OpenDXL, or human-machine teaming that marries machine power with human intellect to achieve better outcomes.
And finally, we think about how we can create new technologies that we recognize are going to be evaded by adversaries when they hit a key point in their life cycle.
Machine learning, deep learning and artificial intelligence are cornerstone technologies that McAfee and much of the industry are building upon, but we must recognize that the adversaries are going to work to innovate around them.
During my MPOWER keynote, I used a machine learning model that is successful in recognizing different handwritten characters, and showed what it might take from a technical perspective to confuse it. The machine learning model initially predicts with 99% probability that the image represented a number “9” character, versus 1% probability that the character is a “4.” By slightly manipulating the pixels of the next the character, probability levels out to 50/50. The image on the right is now at the other end of the spectrum; to you and me it looks like a “9,” but the machine now thinks there’s a 99% chance it’s a “4.”
This same concept can be applied to machine learning capabilities used in cybersecurity defense. We took the same approach and applied it to a malware classifier that judges Android-based malware to be either malicious or benign. By making just slight modifications to the malware, we could fool models into thinking that the code is benign.
Why do I call-out some of the inherent weaknesses in machine learning?
It is because if we close our eyes and disregard that adversaries will attempt something like this, the cyber defense technology that works so well today will fall apart tomorrow. At the same time, if we recognize some of these weaknesses exist, we can put energy into developing defenses today to add resiliency.
This this exactly what we are doing at McAfee. We are looking at all our machine learning capabilities to understand not only how well they work today, but also how they will stand up over time and be resilient and resistant to the evasion attacks of the future.
Objectives, Methods and Innovation
We have to recognize that the adversaries are continuously innovating, and their objectives and methods evolve. They are not focused just on data theft, system breaches, and the sale of stolen information. New business models are driving things like ransomware, where the victim pays the cybercriminal directly, bypassing the risk of reselling data, and monetizing a breach in a very efficient model.
We see things like the weaponization of data, in which attackers can do damage to an individual or an organization by releasing information with the intent to harm them. They are even able to take advantage of changes in the technical ecosystem to find new objectives, such as attacking cloud environments wherein multi-tenant breaches can affect many organizations or users.
Adversaries can take advantage of vulnerabilities by using exploits. They can use stolen credentials to move around environments in such a way that the activity appears to be normal behavior and difficult for defenders to spot.
Sometimes the weakness is not technology. Sometimes it is social, or phishing, or configuration vulnerabilities. Malicious insiders may be authorized actors in an environment.
The Correlation of Detection
Imagine we have a new defense technology that can defend against 5% of the threats on our threat landscape. Should we bring this technology to market when it can stop only 5% of our threats?
You clearly cannot answer that question without more data. If the 5% of threats that this technology can catch is 5% for which existing technologies do not have an answer, such a new technology is very valuable.
This question is not just hypothetical. It is the way that we are engineering and innovating with our new endpoint technology.
McAfee ENS is the most innovative endpoint product on the planet because we have used a set of technologies, each covering a different portion of the threat landscape. You have signature based, you have reputation based, and you have multiple machine learning models. Each technology on its own detects many types of threats, while also leaving some holes.
We must understand what a technology can cover that another technology potentially misses, and how effectively they work together—versus how effectively they work on their own.
Ultimately, part of the answer to out-innovating our adversaries lies in understanding that the correlation of detection technologies is as critical as their efficacy.
My next post will explain how McAfee is understanding correlation as well as efficacy, and how this understanding is paramount to McAfee’s approach to innovation.