Sanjay Kariyappa

AI Research @ NVIDIA

[Sanjay Kariyappa @sanjayatwork]

@trychroma Very interesting work! We’ve also been thinking about model-driven context management in our recent paper SideQuest: https://t.co/Rbgqut3feC

[Sanjay Kariyappa @sanjayatwork]

Link to full paper: https://t.co/vwQPtjVa3p

[Sanjay Kariyappa @sanjayatwork]

Excited to present our #ICML2024 paper 'Progressive Inference' today from 1:30-3 pm! 🌟 We leverage intermediate predictions to provide high-quality input attributions for decoder-only sequence classification models. Come say hi if you're interested in #XAI!

[Sanjay Kariyappa @sanjayatwork]

Come check out our poster if you're attending the conference in person!
Full paper: https://t.co/JeQKMgYFyJ
This work was done during my internship at FAIR/@MetaAI with my amazing collaborators: Chuan Guo, @kiwanmaeng, Wenjie Xiong, G. Edward Suh, @mointweets & @hsienhsinlee.

[Sanjay Kariyappa @sanjayatwork]

Excited to present our #ICML2023 paper Cocktail Party Attack🍸today at 11:00 am in Exhibit Hall 1! We develop a highly scalable attack that leaks private inputs from gradients in federated learning by framing the attack as a blind-source separation problem.

[Sanjay Kariyappa @sanjayatwork]

Our attack works on ImageNet-sized inputs with very large batch sizes (>1000), demonstrating that aggregation alone does not provide meaningful privacy guarantees.

[Sanjay Kariyappa @sanjayatwork]

@SupriyaN20 Looks great! Very clean. Need to try this sometime.

[Gautam Kamath @thegautamkamath]

Very nice to see this work by @sanjayatwork and @mointweets. It's always great to see attacks against methods that are only "intuitively" privacy preserving, like split learning and federated learning (without differential privacy on top).

[Sanjay Kariyappa @sanjayatwork]

@satml_conf This is joint work with @mointweets
Link to paper: https://t.co/2hcisFNaay

[Sanjay Kariyappa @sanjayatwork]

I'll be presenting our work "ExPLoit: Extracting Private Labels in Split Learning" at 1PM ET today at @satml_conf. Our work demonstrates that split learning does not protect label privacy by designing a high-accuracy label leakage attack.

[Sanjay Kariyappa @sanjayatwork]

@fhoussiau Really interesting work! Congratulations!

[Sanjay Kariyappa @sanjayatwork]

How can we steal the functionality of a black-box ML model without using any datasets? Come learn about our data-free attack MAZE being presented at @CVPR between 11:00AM-1:30PM ET. Joint work with @atulprakash and @mointweets.
https://t.co/kuDPAo1i9S

[Atul Prakash @atulprakash]

Our CVPR'21 paper on two novel model stealing attacks: (1) MAZE for completely data-free settings (no datasets needed) and (2) MAZE-PD when limited in-distribution data available. With Sanjay Kariyappa @sanjayatwork and Moin Qureshi @mointweets. https://t.co/fo5yetXHFn

[Sanjay Kariyappa @sanjayatwork]

Excited to present our work on defending against model stealing attacks at #ICLR2021. Stop by poster session 9 at 8 pm ET today for a chat.
Paper: https://t.co/DXZjWtjRGM

[Georgia Tech School of Computer Science @gatech_scs]

@sanjayatwork and @mointweets created a new defense that prevents model cloning. #ICLR2021 https://t.co/4HIjrBVafP

[Sanjay Kariyappa @sanjayatwork]

@mointweets @intel Congratulations! 👏👏