🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Ever wondered how many downloads your model will get on Hugging Face and how fast it grows? How about the number of fine-tuned models over time? Check out our interactive AI model growth predictor to learn more Paper: arxiv.org/abs/2502.15987 Demo: forecasthuggingfacemodels.onrender.com

Giving an oral tmr at 11 am AAAI AI Alignment track for our jailbreak mitigator, Token Highlighter #AAAI2025 We use affirmation loss to find problematic tokens and adopt "soft removal" on them to improve safety w/ Xiaomeng Hu & Tsung-Yi Ho 🔗 shorturl.at/hBJSB

🚨 Is your private data really safe in Federated Learning? Spoiler: Not always. 🚨 Attackers can reconstruct sensitive user data from model updates using gradient inversion attacks. We present CENSOR, a novel defense that breaks these attacks while keeping models utility. 🧵👇

Really honored to receive IEEE Signal Processing Society Industry Professional Leadership Award for my research and industry impact on machine learning robustness and AI safety. Read more about my journey and vision in using signal processing for computational AI safety at x.com/pinyuchenTW/st…

🎉 Congratulations to Dr. Pin-Yu Chen (PhD ECE ’16) on receiving the 2024 Young Professional Industry Leadership Award from the IEEE Signal Processing Society This honor recognizes Chen’s influential contributions to machine learning robustness and AI safety. ⬇️ Read more: bit.ly/4jcmjyZ

🔥Our #ICLR2025 Oral paper "When is Task Vector Provably Effective for Model Editing? A Generalization Analysis of Nonlinear Transformers" will be presented on 04/26, 4:18 p.m. — 4:30 p.m. at Garnet 216-218. Poster pre will be on 04/26, 10:00 a.m. — 12:30 p.m. #341.

🚨 Meet the first 3 invited speakers at #DIGBUG ICML Conference 2025! 🔒 Pin-Yu Chen (IBM): AI safety & robustness 🧠 Sanmi Koyejo (Stanford): Trustworthy ML in health & neuroscience ⚙️ Nouha Dziri (AI2): safety of open LLMs 📅 Submission deadline - May 20: icml2025digbugs.github.io

Your LLM Guard Model is secretly a reliable LLM-finetuning-guardrail! IBM Granite Guardian and LLAMA Guard are particularly suited to tracking harmful levels of fine-tuning data at the token level and making training adjustments during fine-tuning Paper: arxiv.org/abs/2505.17196

A fun project with Ria Vinod Payel Das IBM Research using model reprogramming techniques to "translate" a pretrained English language model for protein sequence representation learning. Our method can be on par with or even better than some of the protein foundation models!

When it comes to AI safety, reinventing the wheel is counterproductive. History is the best teacher in understanding and preventing catastrophic risks in leading technology. Kudos to Kaiyuan Zhang for suggesting that we apply security principles to LLM agents!