Privacy Preserving AI
summary
Claims it is possible to answer questions using data you cannot see by combining PySyft, PyTorch remote execution and cryptography: pointer tensors let code run on hospital servers; a Grid client returns dataset pointers and metadata; .get requires an epsilon privacy budget that adds calibrated noise via differential privacy; .share enables secret‑sharing secure multi‑party computation so models and data are encrypted yet trainable. OpenMined (≈5,000 volunteers) ships prototypes that perform encrypted training and remote prediction and demonstrate cross‑site improvements (example: 96% → 96.1% accuracy) without revealing raw data.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 1:35Access to private data
- 4:43Remote execution
- 6:20Search example data
- 7:40Differential privacy
- 12:33Secure multiparty computation
related talks
45mCrypto-AI: A Guide to Building Private and Decentralized AI Systems
26mDoes using data mean giving up privacy?
38mApplying Privacy-Enhancing Technologies
13mProvably Secure AI for Health Data – Powered by Confidential Computing
19mHow Can We Have Private Access to Intelligence?
29mPrivacy-preserving methods: Building secure projects
29mBalancing Thin Line Between Data Intelligence & Privacy
28mGenerative AI + Proprietary Data