Presented at
DEF CON 31 (2023),
Aug. 11, 2023, 9 a.m.
(240 minutes).
This workshop is a beginner's introduction to deep learning with neural networks, going from fundamentals to the latest in models for image editing, object recognition, and automated pen testing using large language models. It starts with an introduction to the theory behind deep learning, with a few toy examples to give students a feel for how these systems are built. From there we shift focus to a tour of state of the art models with a focus on running open source models locally independent of proprietary corporate systems. These systems include captcha defeat, video search and tracking, and image editing, among others. Finally, students perform a pen testing capstone using AutoGPT and HuggingGPT to understand the latest in emergent large language model reasoning capabilities. Students should have a basic understanding of how to write Python code, the class will build from there. A laptop with 8Gb of RAM and 100GB of free space will be sufficient. Students may bring laptops with more powerful GPUs, but online resources will be available for more GPU intensive models.
Skill Level: Beginner
Prerequisites for students:
- None, this workshop will walk through all steps required to use and apply the models.
Materials or Equipment students will need to bring to participate:
- A laptop with at least 8Gb of RAM and 100GB available hard drive space. Must also be able to run a Linux based VM. This isn't meant to be a high bar, free online resources will be used to supplement their laptop for larger models.
- Students will need an OpenAI API token, which will require setting up a paid account with OpenAI. The final cost for API using in this class should be no more than $5. I wish there was not a requirement for this, but unfortunately some of the cutting edge application I want students to experiment with are only available in high enough quality using OpenAI's products. This may change between this submission and the start date of the class at the rate of current AI advancement.
Presenters:
-
Robert Koehlmoos
Rob works as a lead machine learning engineer focusing on deep learning applications, primarily with language translation. His team works with the full pipeline of training, productionizing, and deploying machine learning applications. He is happy not only talking about theory and research but also the practicalities of model selection and designing products to meet user needs. He previously worked as a data scientist and has strong opinions about effective uses of data visualization and good UI design. He is only a little afraid of AI taking over everything.
Similar Presentations: