Jon Rusert

Research: Attacks/Defenses/Bias in NLP
Forgiven
S.D.G.
"Why do the nations rage and the people plot a vain thing?"
Better Half's art:

[Istra of Glome @byistra]

C.S. Lewis’ version of “go touch grass@

[Kyle Mann @The_Kyle_Mann]

In 2024, Gavin Newsom called on the California legislature to pass laws restricting the use of A.I. to parody politicians. Thankfully, @ADFLegal teamed up with us at @TheBabylonBee to file a lawsuit against these laws. They were ruled unconstitutional earlier this year.

Now, Gavin is parodying politicians with A.I. You're welcome, Governor!

[Jon Rusert @JonRusert]

Last week I presented 2 EMNLP papers (virtually). Here are the recordings and official publications:

RedHerring Attack
Paper: https://t.co/FBV2Z2Ucb2
Video: https://t.co/gKhTahBDxi

Hybrid Select
Paper: https://t.co/fzA8CXkpN4
Video: https://t.co/iyQD03GDWe

Please enjoy!

[Jon Rusert @JonRusert]

Hybrid/Dynamic Select
We propose two new methods to improve efficiency of adversarial attacks. We find both strongly reduce the number of required model queries for successful black-box attacks. We verify the methods on both encoder and LLM models.
Arxiv: https://t.co/EQOW0hY3qj

[Jon Rusert @JonRusert]

Grateful to be a part of two research papers being virtually presented at EMNLP 2025:

1. "RedHerring Attack: Testing the Reliability of Attack Detection"

2. "Overcoming Black-box Attack Inefficiency with Hybrid and Dynamic Select Algorithms"
w/ A. Belde and R. Ramkumar

[Jon Rusert @JonRusert]

RedHerring Attack
We propose/test a novel threat model and attack which causes sows distrust between humans and attack detection models. We expose vulnerabilities in current detection methods as well as investigate defenses to respond.
Arxiv: https://t.co/ErbLMrUxoJ

[Matt Walsh @MattWalshBlog]

There should be social consequences for people who openly celebrate the murder of an innocent man. But there obviously shouldn’t be any legal repercussions for “hate speech,” which is not even a valid or coherent concept. There is no law against saying hateful things, and there shouldn’t be.

[Creeds & Confessions @CreedConfession]

Charlie Kirk
1993-2025

“All death can do to the believer is deliver him to Jesus.”
Quote by John MacArthur

[gabe @allgarbled]

GPT-5 will represent a nearly 25% increase in version number over GPT-4

[Nathan Rusert @NRusert]

Listen to "TreOre "I Thirst" John19:28-29" on PodPoint. https://t.co/OMACUBimr8 via @PodPoint

[Brave @brave]

Just use Brave. 🦁

Brave's built-in protections block ads and trackers across the Web.

Plus, you can enable uBlock Origin directly through our browser (Settings > Extensions > Manifest v2 extensions).

[Jon Rusert @JonRusert]

We propose and explore a black-box attack selection method which aims to be more efficient in finding which words to modify compared to Greedy select (and all its variations [e.g. Importance Score]). We find a good tradeoff between reduction in queries and attack effectiveness.

[Jon Rusert @JonRusert]

📣Excited to announce a paper which will be presented (virtually) at COLING 2025:

"BinarySelect to Improve Accessibility of Black-Box Attack Research"

W/ Shatarupa Ghosh (https://t.co/TjAdDPJT08)

Arxiv:
https://t.co/bBqZn6aUse

[Not the Bee @Not_the_Bee]

We need to seriously evaluate how we let The New York Times beat us to this hilarious satirical headline.

[Jon Rusert @JonRusert]

Video presentation can be found here: https://t.co/dcm3sKDuAQ

[Jon Rusert @JonRusert]

Had a great time presenting my paper VertAttack (https://t.co/ntdCZ1EhFh) at #naacl2024

[Jon Rusert @JonRusert]

We propose and test a new attack which takes advantage of text classifiers' inability to process text vertically. The attack finds high weighted words and rewrites them vertically. This messes up the processing of a classifier, but allows a human to read the obfuscated message.

[Jon Rusert @JonRusert]

🎉Announcing a new paper which was accepted at NAACL 2024 main conference:

"VertAttack: Taking advantage of Text Classifiers' horizontal vision"

Arxiv: https://t.co/hkThwmQ3WX