Thread by @ditrizna

While #DataScience for defensive #infosec is ubiquitous amongst security solutions (generally PE evaluation, but system telemetry analysis is catching up), #MachineLearning for #offensive security is still in the discovery phase.

A brief review on promising techniques in🧵

1) Adversarial malware - use AI to generate evasive malware samples. A lot of publications here. I assume some AI-powered packers might appear in the next 2 to 4 years. Good starting point if you want to dig deeper here is this library by @zangobot: github.com/pralab/secml_m…

2) Another branch of research is using **generative** models for some offensive needs like (a) ML-powered fuzzing, e.g. arxiv.org/abs/1908.01262, but search for more work, (b) password list generators, e.g. arxiv.org/abs/1709.00440, (c) C&C needs (bwall.github.io/markov-chains-…).

3) Uniquely stands Sandbox evasion with ML by Pearce et al. (@moo_hax): arxiv.org/abs/2007.06763

4) SAST/DAST analysis with NLP-like models -- sorry, I am not proficient in this area to give valuable references, but it is a highly promising direction with existing prototype (or even commercial) solutions.

These were conventional #offensive security tasks. Separately should be considered:

5) Security of AI applications itself, adversarial machine learning. This is an unsolved problem, with the high interest from the research community and funding. Refer: atlas.mitre.org