Machine Learning over Encrypted Data with Fully Homomorphic Encryption
summary
Claims that Fully Homomorphic Encryption lets models run on encrypted inputs and shows a usable toolchain: an open source library Concrete-ML that compiles PyTorch/Scikit-learn/XGBoost models to an FHE runtime using TFHE and programmable bootstrapping (PBS) for activations. Quantization (3–16 bits) and pruning convert weights to integers; tree ensembles and linear models run with little loss, trees hitting near-floating accuracy at six bits and sub‑2s on a spam test; CIFAR models and image/filter and sentiment demos run end‑to‑end from GitHub examples.
AI-written summary · about
chapters (12)
- 0:00Introduction to speakers
- 1:48Fully homomorphic encryption
- 6:16Hugging Face demonstration
- 9:48How FHE works
- 14:06Machine learning with FHE
- 18:48Concrete ML library
- 21:42Tree-based models
- 26:47Linear models
- 28:11Deep learning models
- 32:34Custom model training
- 33:37Project examples
- 38:26Q&A session
related talks
37mOpen-source Machine Learning on Encrypted Data
40mTrust No One
44mIP Protection and Privacy in LLM: Leveraging Fully Homomorphic Encryption
33mThe Promise of Trustworthy AI
06mLightning Talk: Multi-Party Learning Using Homomorphic Encryption with Kubeflow...
38mPrivacy in the Open Source: Homomorphic Encryption
54mFully Homomorphic Encryption
23mBootstrapping Homomorphic Encryption via Functional Encryption