📣 Our work Universal Train-Time Poison Defense method, PureGEN with @HessianFree @Jimmeryy @alexrbranch, recently got accepted into #neurips2024! I'm excited to share a Part 2 blog post discussion on how to train Energy-Based Models and Diffusion Models for image poison defense

Github Repo: https://t.co/NhJILzxG7z HuggingFace Pre-Trained Models:

👍Conclusion Training custom EBMs and DDPMs form the core of PureGen’s approach to robust poison defense. By leveraging the dynamics of generative models, PureGen effectively purifies data, removing adversarial perturbations while preserving essential features Check out the

❔Intuition Behind Purification Using EBMs and DDPMs The purification process with both EBMs and DDPMs relies on their capacity to “push” images back to the clean data manifold. So, by using Langevin dynamics of EBMs and DDPMs, we transport poisoned images back toward the low

🎞️Training Diffusion Models (DDPMs) for Purification DDPMs operate by iteratively adding noise to images in a forward process then we learn to iteratively subtract noise in a reverse process to restore an image from the original data distribution. In PureGen-DDPM, the forward

🔋Training Energy-Based Models (EBMs) for Purification EBMs model the data distribution explicitly through an energy function, assigning low energy values to realistic (clean) samples and high values to improbable (poisoned) samples. We can then use Markov Chain Monte Carlo

💡PureGen Pipeline and Motivation (Recap) As a quick recap, PureGen is a universal data purification framework that leverages the dynamics of specifically Energy-Based Models (EBM) and Denoising Diffusion Probabilistic Models (DDPM). This preprocessing pipeline is designed to be

Riding the physics wave... In our latest #NeurIPS2024 paper PureGen with @HigherSNR @Jimmeryy @alexrbranch we use Energy Based Models to purify adversarial poisons. We do this by initializing the MCMC sampling chain from an image and reduce its energy down a manifold of the

🚀Super excited to report that PureGEN was accepted to #neurips2024 We provide near complete protection against SoTA poisons with little to no degradation in natural accuracy. None of this would have been possible without the amazing team @HigherSNR @Jimmeryy @alexrbranch

I just published PureGen: Deep Learning Poison Attack Defense via Generative Models https://t.co/WWNdo0nbb2 High-level details on our paper: PureGen: Universal Data Purification for Train-Time Poison Defense via Generative Model Dynamics: https://t.co/LNluWrpodP

📢 Excited to introduce PureGEN: Universal Data Purification for Train-Time Poison Defense! This method ensures robust ML models by purifying datasets with minimal impact on generalization. https://t.co/WfBgNZd1Tj https://t.co/nTICmrPDPI