AI and cybersecurity researcher @robustintelligence. Formerly, head of AI @Sophos. Views are my own.
Washington, DC
Joined April 2007
- Tweets194
- Following266
- Followers341
- Likes419
Konstantin Berlin retweeted
First talk of the session: "Web content filtering through knowledge distillation of Large Language Models" by Tamas Voros!
Konstantin Berlin retweeted
Reminder to submit your full length paper or extended abstract by Friday, August 4th to be considered for this year's conference! camlis.org/call-for-papers
#CAMLIS2023 #LLMs #Analytics #Statistics #DataScience #infosec
Konstantin Berlin retweeted
Why is this a concern about open-source models rather than language models in general?
Check out a quick writeup about our work on LLM for web classification.
news.sophos.com/en-us/2023/0…
Konstantin Berlin retweeted
For more, see the full report, which links to the SophosAI paper. news.sophos.com/en-us/2023/0…
Konstantin Berlin retweeted
The results were models that were 175 times smaller than the LLMS and just as accurate. While that accuracy (using just URLs) was not optimal, adding additional data (such as HTML from the page) and allowing multiple labels could raise it significantly.
Konstantin Berlin retweeted
Sophos X-Ops' AI team has been working on a way to apply the power of Large Language Models (LLMs) like GPT to security tasks - but without the massive expense of having to deploy them. Instead, LLMs can be used to train smaller models. It’s called knowledge distillation.
Creating small/accurate ML models for URL webcat is difficult because of the long tail issue. Excited to share our work with @tamasV2 and @SPBergeron on using #LLM and #GPT3 that significantly moves the needle.
arxiv.org/abs/2305.05027
@Sophos @SophosXOps
I feel that one way to avoid an AI apocalypse would be to not let GPT6 launch nuclear weapons without additional guardrails. Let it try to destroy humanity while running on my laptop, receiving only cooking recipe requests as input.
Konstantin Berlin retweeted
More CAMLIS? More CAMLIS!
The Conference on Applied Machine Learning for Information Security is returning to Arlington VA, October 19-20, 2023! Details at camlis.org/ (where you can join our slack). Apply to our call for volunteers here: docs.google.com/forms/d/e/1F…
Konstantin Berlin retweeted
Sophos X-Ops has been testing three prototype projects that demonstrate the potential for GPT-3.5—the algorithmic brain of #ChatGPT—to be used for good in infosec. All three were developed using the few-shot learning method to reduce the need to collect large volumes of data. /1
A nice writeup of some #ML #GPT work we have been doing over the past year at @Sophos @SophosXOps
news.sophos.com/en-us/2023/0…
Konstantin Berlin retweeted
By telling scientists they must publish, you get:
1. higher-quality research, more reliable results, less self-delusion
2. better scientists whose reputation will flourish
3. easier external collaborations
4. better research evaluation
5. better internal impact
6. prestige
Good alert prioritization inside SOCs can make the difference between discovery of an attack and just paying a lot of money for an ineffectual team. Excited to share our AutoML approach to this. @tfq_salma @gelman_ben @tamasV2 @SophosXOps @Sophos
arxiv.org/abs/2302.06648
Konstantin Berlin retweeted
The finished version of our SaTML'23 position paper is now (finally!) on arXiv: arxiv.org/abs/2212.14315
Thanks again to: @drhyrum @savinodam @dfreeman02 @fbpierazzi @AlejandroRoundy! (but also to @dagstuhl!!)
@satml_conf
Some more work by @younghoo_au and @joshua_saxe on training #GP3 models to translate natural language requests to vendor XDR/MDR queries. Now a complex task hidden behind complex interfaces and database schemas. @SophosAI @SophosXOps @Sophos
ai.sophos.com/2022/12/15/nat…
With all the recent talk about #ChatGPT we at @SophosAI have been busy thinking about practical applications of LLMs to cybersecurity. Younghoo Lee wrote a blog post about some things that we are exploring ai.sophos.com/2022/12/15/gpt… @SophosXOps @Sophos
Konstantin Berlin retweeted
Don't miss out on #NLP the #Cybersecurity Panel at #ODSCWest! Join Adarsh Kyadige, @kberlin @joshaaayyyyy, George Williams, Nick Gregory as they dive into the cutting-edge of language models for practical security use cases! Join us 🤖🤖🤖🤖🤖🤖🤖🤖🤖🤖🤖 hubs.li/Q01r95dB0
