Page 85 - Cyber Defense Magazine for August 2020

The Rising Cost of Failure

How much is cybersecurity costing companies? In a recent Ponemon-Deep Instinct survey of IT and IT
security practitioners, only 40% of respondents believed their budgets were sufficient for achieving a
robust cybersecurity posture.

These budgets are predominantly funneled into containing and remediating threats rather than preventing
them – in large part because cyber staff are overwhelmed with the amount of data that they need to
monitor. Yet this “assume a breach and then contain” approach comes at a big cost, with the time and
money spend remediating attacks costing well into the hundreds of thousands of dollars. The value of
preventing a cyber-attack ranges from $400,000 to $1.4 million, depending on the nature of the attack. If
an attack is the first of its kind, it’s virtually guaranteed to succeed with absent strong preventative
capabilities, and organizations stand to lose upwards of $1 million per successful attack.

Subpar Solutions, Subpar Results

Why are current approaches to cybersecurity proving so inadequate? Because they over-rely on the
human intervention.

Specifically, most AI-based cybersecurity solutions are powered by traditional machine learning (ML),
which is inhibited by a number of limitations that have become substantial problems in the recent past.
Chief among these limitations is data: ML models are trained on only a fraction of the available raw data,
and are trained on features identified by experts.

Human error, of course, also comes into play, even when highly specialized computer scientists with
expertise in cybersecurity carry out ML feature engineering. These professionals excel at training ML
models on known threats – but even seasoned cybersecurity professionals are unable to anticipate
emerging, first-seen attacks, that are designed to be evasive. Hackers of course, understand this, which
is why they now building malware that is capable of fooling ML models into classifying it as benign.

Finally, there’s a limit to the size of the dataset for training ML systems before reaching learning curve
saturation – the point past which the system no longer improves its accuracy.

Given these limitations, ML systems struggle to detect new, previously unseen malware, while generating
high rates of false positives. Just as the cost of an unprevented attack can deliver a real blow to the
bottom line, the time and resources required to investigate false positives also strains security teams’
resources. This, in turn, breeds a sense of “alert fatigue,” making teams more prone to error when
genuine threats emerge.

Simply put, AI trade-offs – not understaffed cybersecurity teams – may be one of the biggest inhibitors to
achieving a resilient cybersecurity posture.

Cyber Defense eMagazine – August 2020 Edition 85
Copyright © 2020, Cyber Defense Magazine. All rights reserved worldwide.