Share
Embed
When Data Collaboration Meets Privacy: Privacy-enhancing Technologies on AWS
Presented at AWS Community Day Bay Area Security Edition 2024
Richard Fan
March 21, 2024
Transcript
44 slides
-
Who am I? Richard Fan AWS Security Hero Security Engineer
AWS Nitro Enclaves contributor -
Privacy-enhancing Technologies (PETs) • Homomorphic encryption (HE) • Trusted Execution
Environments (TEE) • Secure multi-party computation (SMPC) • Differential Privacy (DP) • Federated learning (FL) -
Trusted Execution Environment (TEE) Secured area of a processor Isolated
from the rest of the system Cannot accessed/modified by system admins or OS E.g. AWS Nitro Enclaves, Intel SGX, AMD SEV-SNP, Apple Secure Enclave, Arm TrustZone, NVIDIA Confidential Computing -
AWS Nitro Enclaves Isolated virtual machine Communication over a secure
channel No admin access No persistent storage No external networking https://aws.amazon.com/blogs/aws/aws-nitro-enclaves-isolated-ec2-environments-to-process-confidential-data/ -
AWS Nitro Enclaves Prevent unauthorized access Protect data in use
Ensure application integrity -
Attestation Document Describe the running enclave Signed by AWS Provide
root of trust from AWS -
Federated Learning Machine Learning Training data stay locally Central node
aggregate models https://royalsociety.org/-/media/policy/projects/privacy-enhancing-technologies/From-Privacy-to-Partnership.pdf -
Federated Learning – Use cases Medical ML models COINSTAC –
FL tools on Neuroimages Estimate brain age from MRIs Accuracy is similar to tradition models https://www.biorxiv.org/content/10.1101/2021.05.10.443469v2.full.pdf -
Amazon SageMaker – Distributed Training Run training jobs in parallel
Scale across instances Make training fast https://aws.amazon.com/campaigns/sagemaker/ -
Amazon SageMaker – Federated Learning Connect accounts via VPC Peering
Federated Learning Libraries: • FedML • OpenFL • Flower • TensorFlow Federated https://aws.amazon.com/blogs/machine-learning/machine-learning-with-decentralized-training-data-using-federated-learning-on-amazon-sagemaker/ https://aws.amazon.com/blogs/machine-learning/part-1-federated-learning-on-aws-with-fedml-health-analytics-without-sharing-sensitive-data/ https://aws.amazon.com/blogs/machine-learning/enable-data-sharing-through-federated-learning-a-policy-approach-for-chief-digital-officers/ -
Revealing medical records of Governor, 1997 Ethnicity Visit date Diagnosis
Procedure Medication Total charge Birth date Gender ZIP Name Address Date registered Affiliation Date last vote Voter registry Medical record https://epic.org/wp-content/uploads/privacy/reidentification/Sweeney_Article.pdf -
Birth date Gender ZIP Revealing medical records of Governor, 1997
Ethnicity Visit date Diagnosis Procedure Medication Total charge 6 people Medical record Voter registry Name Address Date registered Affiliation Date last vote https://epic.org/wp-content/uploads/privacy/reidentification/Sweeney_Article.pdf -
Revealing medical records of Governor, 1997 Ethnicity Visit date Diagnosis
Procedure Medication Total charge 3 people Medical record Voter registry Birth date Gender ZIP Name Address Date registered Affiliation Date last vote https://epic.org/wp-content/uploads/privacy/reidentification/Sweeney_Article.pdf -
Revealing medical records of Governor, 1997 Ethnicity Visit date Diagnosis
Procedure Medication Total charge 1 person Medical record Voter registry Birth date Gender ZIP Name Address Date registered Affiliation Date last vote https://epic.org/wp-content/uploads/privacy/reidentification/Sweeney_Article.pdf -
De-anonymization on Netflix Prize Dataset, 2006 Subscribe #12345 Catch Me
If You Can Forrest Gump Fahrenheit 9/11 Jesus of Nazareth John Catch Me If You Can Forrest Gump https://arxiv.org/pdf/cs/0610105.pdf -
De-anonymization on Netflix Prize Dataset, 2006 John Catch Me If
You Can Forrest Gump Fahrenheit 9/11 Jesus of Nazareth John Catch Me If You Can Forrest Gump Fahrenheit 9/11 Jesus of Nazareth https://arxiv.org/pdf/cs/0610105.pdf -
No person is exactly the same as the others All
data is potential PII -
Differential Privacy (DP) A statistical measure How much an individual
impact a dataset Important concepts: • Privacy Budget (ε) • Noise -
Lower ε (Leak less privacy) Name Country City Alice US
LA … … … Bob US Chicago Peter UK London Privacy Budget (ε) – How best John can hide from my question? How many people live in US? Name Country City John US Monterey Alice US LA … … … Bob US Chicago Peter UK London 100 99 -
Higher ε (Leak more privacy) Name Country City Alice US
LA … … … Bob US Chicago Peter UK London Privacy Budget (ε) – How best John can hide from my question? How many people live in Monterey? Name Country City John US Monterey Alice US LA … … … Bob US Chicago Peter UK London 1 0 -
Name Country City John US Monterey Alice US LA Bob
US Chicago … … Peter UK London Noise – Helping John hide How many people live in Monterey? 0? 1? 2? 10? -
Differential Privacy (DP) Lower ε → More privacy preserved Higher
noise → Lower ε → More privacy preserved Higher noise → Lower accuracy → Lower usability -
AWS Clean Rooms Sharing Glue Tables across accounts Analysis Rules
limit query capability Differential Privacy (In Preview) -
AWS Clean Rooms – Demo Member ID City Gender Education
… … … … 123456 Charlottetown Male Doctor … … … … … … … … Loyalty member database -
Recap Trusted Execution Environment (TEE) → AWS Nitro Enclaves Federated
Learning (FL) → Amazon SageMaker Differential Privacy (DP) → AWS Clean Rooms