Machine Learning Failures

Full post for this episode (https://cisoseries.com/defense-in-depth-machine-learning-failures/)

Is garbage in, garbage out the reason for machine learning failures? Or is there more to the equation?

Check out this post and discussion for the basis of our conversation on this week’s episode co-hosted by me, David Spark (@dspark), the creator of CISO Series and Allan Alford (@AllanAlfordinTX), CISO at Mitel. Our guest for this episode is Davi Ottenheimer (@daviottenheimer), product security for MongoDB.

Thanks to this week’s podcast sponsor, Remediant

81% of cyberattacks utilize stolen administrative credentials. Yet, legacy enterprise password vaults solve only a fraction of the problem and are difficult to rollout. Remediant's SecureONE takes a new approach to privileged access management: offering agent-less, vault-less, continuous detection and just-in-time-administration. Learn what Remediant can do in a half-day POC deployment.

On this episode of Defense in Depth, you'll learn:

• Don't fall victim to believing that success and failure of machine learning is isolated to just garbage in/garbage out. It's far more nuanced than that. Some human actually has to determine what is considered garbage in and what is not.
• It only takes a very small amount of data to completely corrupt and ruin machine learning data.
• This knowledge of small infection can spread and corrupt all of the data and can have political and economic motivations to do just that.
• We have failures in human intervention. Machine learning can just magnify that at rapid rates.
• While there are many warning signs that machine learning can fail, and we have the examples to back it up, many argue that competitive environments don't allow us to ignore it. We're in a use it or lose it scenario. Even when you're aware of the pitfalls, you may have no choice but to utilize machine learning to accelerate development and/or innovation.