Page 16 - Cyber Defense Magazine RSA Edition for 2021

legitimate activity may be blocked and disrupt business activity, or alternatively malicious activity may be
mistakenly identified as being OK and an alert not generated, leaving the organization open to risk.

Hurdles AI/ML detection tools must overcome

Despite their potential, AI/ML detection tools are not a panacea. They can miss threats and they can flag
activities as malicious when they’re not. In order for these tools to be trusted, the alerts they raise and
the decisions they make need to be continually validated for accuracy.

If we train our AI and machine learning algorithms the wrong way, they risk creating even more noise,
and alerting on things that are not real threats (“false positives”). This wastes analysts’ time as they chase
these phantoms unnecessarily, only to discover they’re not real threats. Alternatively, they may also miss
threats completely that should have been alerted on (“false negatives”).

How do we guard against the pitfalls?

In the case of false positives, we need to validate that a detected event is not malicious and train the tool
to ignore these situations in future - while at the same time ensuring this doesn’t cause the tool not to
alert on similar issues that are in fact malicious. The key to doing this effectively is having access to
evidence that enables accurate and timely investigation of detected threats. Recorded packet history is
an indispensable resource in this process - allowing analysts to determine precisely what happened and
to accurately validate whether an identified threat is real or not.

Dealing with false negatives is more difficult. How can we determine whether a threat was missed that
should have been detected? There are two main approaches to this. The first is to implement regular,
proactive threat hunting to identify whether there are real threats that your detection tools - including
AI/ML tools - are not detecting. Ultimately, threat hunting is a good habit to get into anyway, and if
something is found that your AI/ML tool missed the first time around, it provides an opportunity to train it
to correctly identify similar threats the next time.

The second approach is using simulation testing - of which one example is penetration testing. By
creating simulated threats, companies can clearly see if their AI/ML threat detection tools are identifying
them correctly or not. If they’re not, it’s once again an opportunity to train the tool to identify similar activity
as a threat in future.

16