Very interesting insights: Image AutoRegressive Models might be better and faster than Diffusion models BUT they leak more #privacy!
If you want to learn more about these trade-offs, check out our latest insights!
🚨 Image AutoRegressive Models Leak More Training Data Than Diffusion Models🚨
IARs — like the #NeurIPS2024 Best Paper — now lead in AI image generation. But at what risk?
IARs:
🔍 Are more likely than DMs to reveal training data
🖼️ Leak entire training images verbatim
🧵 1/
A truly great outcome for a summer internship in our SprintML lab. If you are interested in working on these kinds of cool projects with us this summer, please apply:
🖇️ https://t.co/6QXEb7wCVG
🔥 New ICLR 2025 Paper!
It would be cool to control the content of text generated by diffusion models with less than 1% of parameters, right?
And how about doing it across diverse architectures and within various applications? 🚀
🫡 Together with @lukxst, we show how:
🧵 1/
Have you ever wondered how you can train an #ML#model with #privacy guarantees? In our lecture on #DifferentiallyPrivate ML, we cover the two canonical algorithms: #DPSGD and the #PATE. We discuss principles, privacy analysis and accounting, and effects: https://t.co/xpAPMGSz2k
Did you know that the predictions of an #ML model can give away #private information of the model’s training data? We cover the problem of #privacy leakage in our new lecture on #trustworthy ML and introduce you to #DifferentialPrivacy: https://t.co/woHWxa1clu
Happy to share your recently accepted #ICLR2024 papers on #arxiv? But spending hours and hours on cleaning off comments, compiling, re-compiling?
You don't have to - just follow the simple guide we compiled for our #PhD#students 👇
https://t.co/ilzQgkqb4A
Check @adam_dziedzic's latest talk on differential privacy for prompting LLMs: https://t.co/QqNsoDBziw It covers our latest paper from #NeurIPS2023 https://t.co/2pM212NzPx We propose to privately learn to prompt and show how to create private discrete and continuous/soft prompts
@adam_dziedzic and @fraboeni are shortly joining the panel on backdoors in #ML at the BUGS workshop @NeurIPSConf - pass by in room 205 (More info: https://t.co/sNI9zdPn5V) 🎤🎤
🔓Eager for some Privacy-Preserving Machine Learning before Christmas? 🎄
🎓Join us on Dec. 21 for double research seminar with @fraboeni and @adam_dziedzic of @CISPA@SprintML
https://t.co/UDkDBmH8uS
We propose the first peer-to-peer (P2P) learning scheme that is secure against malicious servers and robust to malicious clients. Our generic framework transforms any algorithm for robust aggregation of model updates to tackle these vulnerabilities: https://t.co/ghspayjcm5