Congrats to PhD students Ali Naseh and Kalpesh Krishna, along with Associate Professors Mohit Iyyer and Amir @Houmansadr, for winning a Distinguished Paper Award for “Stealing the Decoding Algorithms of Language Models” at ACM CCS 2024 2023! 🙌 Read the paper: brnw.ch/21wFg5H

#NeurIPS2023 here we go, we are organizing a cool BUGS workshop on backdoors with a super interesting talks and panel with leading researchers

neurips2023-bugs.github.io

I am looking for PhD students to work together on privacy and security problems in “AI Systems”. We will focus on language models, agents, ML services, and study where they fail and how to make them work better. Apply by December 15. #phdlife #manningcics

No more questions on how domain expertise gets papers into conferences.

The PhD Applicant Support Program invites all #ManningCICS PhD applicants to a live panel webinar on Nov. 9. Applicants who identify as an underrepresented minority in STEM are also invited to receive one-on-one mentoring – Sign up by Nov. 20. Learn more: paspumasscs.github.io

Now I know why a lot of spam went through to my mailbox and why this spam had PDFs attached-> attackers now embed phishing links into PDF which is hard to check.

“From Attachments to SEO: Click Here to Learn More about Clickbait PDFs!” by Stivala et al.

arxiv.org/abs/2308.01273

Cobgrats to CISPA folks!

🕷Neural nets have dead ReLUs that we prune or ignore.

🎃But Halloween blurs the boundary between living and dead. Tonight, these ReLUs cannot be ignored.🎃

This is a thread for those brave enough to learn of the CLIP model’s death neuron. Scroll away if you must sleep tonight.

idk, if this dataset gets released, it might be possible to reidentify people who contributed to the dataset, not sure this is appropriate.

Incredibly inspirational academic!

great to see @AbdelnabiSahar quoted by WIRED

I am stoked to be featured on the cover of this well-written NYT article nytimes.com/2023/09/25/tec…

I believe *alignment* is the secret sauce behind ChatGPT. Having worked on RLHF, including data collection from external vendors, and finetuning hundreds of open-access models at…

interesting observation, possibly positive implications for retrieving private data

amazing paper by Umar Iqbal et al, clear and nice framework to explore vulnerabilities in LLM plugins: arxiv.org/abs/2309.10254

Please consider submitting! exciting set of speakers and very timely to discuss backdoor attacks and watermarking problems!

ML is always deployed inside a system adding more attack vectors