Protecting Sensitive Data and Models for Machine Learning
summary
Claims production-grade confidential ML by running Python functions on encrypted inputs so both model and data remain encrypted during processing. Platform uses AWS Nitro Enclaves with attestation, envelope encryption (AES+RSA), a CLI and Python SDK that deploys an EIF containing an ONNX ResNet50 and returns top-5 classes. Demo shows building a deploy folder, deploying a 150MB function, encrypting bytes client-side and invoking the function (token hiccup in demo but flow is reproducible). Libraries include serdio helpers for numpy byte conversion and fine-grain decryption inside enclaves.
AI-written summary · about
chapters (7)
- 0:00Introduction to talk
- 3:14What is confidential Computing
- 8:10Why use enclaves for ML
- 15:33Cape platform overview
- 22:34Confidential image classification demo
- 35:17Extra Python SDK features
- 39:51Q&A and resources
related talks
21mProtecting Sensitive Data and Machine Learning Models
23mProtecting Sensitive Data in Web Apps
37mOpen-source Machine Learning on Encrypted Data
26mMachine Learning on Encrypted Data with CrypTen
44mPrivacy by Design Machine Learning
47mTowards Confidential AI for the Masses!
21mEnd-to-end secure ML Development
28mGenerative AI + Proprietary Data