The presence of “adversarial change” in machine learning (ML) training data seems to be the differentiator between ML being useful for finding bad things vs. ML being useful for spotting things that are, as an example, very likely *good* based on their consistency and prevalence.

Keith

@kwm

Anyone interested in the security applications of machine learning (ML) should watch this talk by @halvarflake:

Video: https://youtu.be/BWFdxAG_TGk

Slides: https://2017.zeronights.org/wp-content/uploads/materials/ZN17_Thomas_Dullien_Machine%20learning,%20offense,%20and%20the%20future%20of%20automation.pdf

The section ~19:10 is a useful grounding in reality re: ML as a means of accurate threat detection.

9/10/2020, 7:50:00 AM

Favs: 22

Retweets: 6

link

Keith

@kwm

The presence of “adversarial change” in machine learning (ML) training data seems to be the differentiator between ML being useful for finding bad things vs. ML being useful for spotting things that are, as an example, very likely *good* based on their consistency and prevalence.

9/10/2020, 7:59:40 AM

Favs: 1

Retweets: 0

link

Keith

@kwm

“It’s interesting that the place where we put the most effort in using machine learning is also the place that’s least suited for it.”

—@halvarflake, "Machine learning, offense, and the future of automation”

Video: https://www.youtube.com/watch?v=BWFdxAG_TGk (~22:30)

9/10/2020, 8:17:22 AM

Favs: 2

Retweets: 0

link

Keith

@kwm

@halvarflake This video has since been made private, but it's archived here: https://web.archive.org/web/20200918080551/https://www.youtube.com/watch?v=BWFdxAG_TGk

7/3/2023, 9:01:02 AM

Favs: 2

Retweets: 0

link