A European bank is developing a fraud-detection model using transaction-level data that include customer identifiers, account numbers, and purchase histories. Because the data are subject to both the GDPR and the U.S. Gramm-Leach-Bliley Act (GLBA), no personal data may be transferred outside the EU. At the same time, a U.S. data science team needs record-level information to engineer new features and run offline experiments. Which data-handling strategy best satisfies the regulatory constraints while still giving the U.S. team the analytical flexibility it needs?
Hash all direct identifiers with a salted SHA-256 function and export the resulting pseudonymized table to the U.S. team.
Train a differentially private generative model on the EU dataset and share only the synthetic transactions it produces.
Encrypt the complete original dataset with AES-256, send it to the U.S. analysts, and allow them to decrypt it inside an isolated virtual private cloud.
Add Gaussian noise to numeric fields and round timestamps to the nearest hour before transferring the modified dataset.
Generating a differentially private synthetic dataset is the best strategy. This approach involves training a privacy-guaranteed generative model within the EU and then sharing only the artificial records it produces. Because these synthetic records are not linked to any real individual, they fall outside the scope of both GDPR 'personal data' and GLBA 'non-public personal information', making cross-border transfer permissible. Hashing direct identifiers constitutes pseudonymization, not anonymization; under GDPR, such data is still considered personal and its export is restricted. Encrypting the original data for transfer and decrypting it in the United States would be a direct violation of the data transfer ban, as the full PII would be processed in the U.S. Simply adding random noise and rounding timestamps does not guarantee against re-identification and is insufficient to declassify the data as personal information under GDPR's strict standards.
Ask Bash
Bash is our AI bot, trained to help you pass your exam. AI Generated Content may display inaccurate information, always double-check anything important.
What is a differentially private generative model?
Open an interactive chat with Bash
Why is pseudonymization different from anonymization under GDPR?
Open an interactive chat with Bash
What makes synthetic data transfer permissible under GDPR and GLBA?