BastionAI: Towards an Easy-to-use Privacy-preserving Deep Learning Framework
summary
Claims an easy-to-use privacy-preserving deep learning framework that runs models inside secure enclaves + Differential Privacy and reports concrete results: fine-tuned BERT on the SMS Spam Collection Data Set within an enclave, with a stated privacy budget and a reproducible pipeline. Benchmarks show reduced communication and, in the presented experiments, about 10× faster than a federated DP baseline for similar privacy settings. Code and deployment tools live at https://github.com/mithril-security/bastionai.
AI-written summary · about
chapters (8)
- 0:00BastionAI introduction
- 3:34Federated learning challenges
- 9:15Confidential computing explained
- 12:34Secure enclave attestation demo
- 15:46Fortified learning architecture
- 20:21Performance benchmarks
- 22:30Cloud providers for confidential computing
- 23:58Q&A
related talks
31mScaling privacy embedded distributed ML systems
14mEstablishing Privacy Metrics for Genomic Data Analysis
29mBalancing Thin Line Between Data Intelligence & Privacy
44mPrivacy by Design Machine Learning
01h 01mToward Provably Private Federated Learning
21mTowards Building a Responsible Data Economy
27mCreate Value and Insight Through Privacy-preserving Data Collaboration [...]
16mDP-BREM: Differentially-Private and Byzantine-Robust Federated Learning...