A Model for Human and Machine Interaction: Human-Machine Teaming Grows up

Security operation centers (SOCs) are struggling to keep up with attackers, and artificial intelligence (AI) has failed to deliver significant improvements. The industry has been successful at applying AI to malware detection and user and entity behavior analytics (UEBA) using deep neural networks and anomaly detection. But other core SOC jobs such as monitoring, triage, scoping, and remediation remain highly manual. Some repetitive and low-value tasks can be assisted with automation, but tasks that require analysis and creativity are hard to capture in code. Even worse: Imagine trying to automate the investigation of an undiscovered attack technique.

Automation and current AI solutions depend upon a human observing and understanding a threat, then building a model or writing code. The time gap between the human observing a phenomenon and the machine helping is the reason why attackers often have the upper hand. In order to get ahead, we need to make AI systems learn and interact directly with practitioners at the SOC.

The idea behind human-machine teaming (HMT see [1] and [2]) is to put the human in the AI algorithm loop. In a SOC context, the human has the intuition to find a new attack technique and the creativity to investigate it using company tools. Using human input, the machine gathers information and presents it back in a summary to manage the human cognitive workload. As a result of the human-machine interaction, the machine learns to better proceed in new scenarios, while the human continues to adapt, focusing on higher-value tasks.

Several products put the human in the loop, but few empower the human to perform high-order cognitive tasks.

Research shows that unsupervised anomaly detection can be improved by asking the human to examine alerts when classification confidence is low. This approach improves detection by 4X and reduces false positives by 5X [3]. More importantly, the system teaches itself to address adversaries’ changing tactics.

Our assessment of the current SOC tools landscape shows that several products put the human in the loop, but very few empower the human to perform high-order cognitive tasks. In order to understand where we stand as an industry and what the gap is, we clustered tools into four groups.

Most cybersecurity products today deliver HMT1 and HMT2 capabilities. McAfee Investigator delivers HMT3 and our engineers are working toward HMT4.

On the vertical axis, we have ascending levels of cognitive tasks that humans bring to the team, while on the horizontal axis we have machine capabilities. An assumption of this model is that a human is not able to exercise high-order tasks if she also has to perform low-level functions. This is similar to a Maslow pyramid psychology model. As the machine starts to interact with the human at a higher level of cognition, the team becomes more effective and the degree of human-machine teaming increases from HMT0 to HMT4.

Most of the products in the industry today revolve around the first two iterations of human-machine teaming, known as HMT1 and HMT2. In these scenarios, humans interact with products by analyzing data and providing explicit orders on how to drill down and gather additional data. In some products, humans are able to elevate their work by getting insights and applying their intuition and context to them.

What is clearly missing are products that can take directional feedback, for instance: “Get me evidence that supports potential lateral movement on this case”. We are also missing products that can learn by  observing the human at work, for instance, learning to dismiss the alerts that humans have investigated and dismissed in the past.

At McAfee we are using this HMT maturity model as a guide to building better features and tools for the SOC. We recently launched McAfee Investigator [4] to help triage alerts faster and more effectively. Investigator, which uses a question answering approach to leverage expert knowledge [5], can take directional feedback from the human to pivot an investigation (HMT3). Our goal is to develop Investigator to a point where it can learn directly from practitioners (HMT4).

Learn more about human-machine teaming here.

 

[1] S. Grobman, “Why Human-Machine Teaming Will Lead to Better Security Outcomes,” 13 July 2013. [Online]. Available: https://securingtomorrow.mcafee.com/executive-perspectives/human-machine-teaming-will-lead-better-security-outcomes/
[2] B. Kay, “News from Black Hat: Humans Collaborate and Team with Machines to Work Smarter,” 25 July 2017. [Online]. Available: https://securingtomorrow.mcafee.com/business/news-black-hat-humans-collaborate-team-machines-work-smarter/
[3] K. Veeramachaneni, I. Arnaldo and V. Korrapati, “AI^2 : Training a big data machine to defend,” IEEE 2nd International Conference on Big Data Security on Cloud, 2016.
[4] “McAfee Investigator,” [Online]. Available: https://www.mcafee.com/us/products/investigator.aspx
[5] F. M. Cuenca-Acuna and I. Valenzuela, “The Need for Investigation Playbooks at the SOC,” 2017. [Online]. Available: https://www.sans.org/summit-archives/file/summit-archive-1496695240.pdf   
McAfee does not control or audit third-party benchmark data or the websites referenced in this document. You should visit the referenced website and confirm whether referenced data is accurate.
McAfee technologies’ features and benefits depend on system configuration and may require enabled hardware, software, or service activation. Learn more at mcafee.com. No computer system can be absolutely secure.

Leave a Comment

seven + 9 =