DP-NMT: Scalable Differentially-Private Machine Translation

DP-NMT investigates scalable differentially-private training for neural machine translation. This work adapts and evaluates DP training approaches across multiple NLP model architectures and benchmark datasets (WMT16, BSD).

Contributions:

• Adapted differentially-private training to multiple NLP models and translation datasets
• Resolved out-of-memory (OOM) issues within the JAX training framework
• Extended the training pipeline to support additional datasets beyond the original scope

Role: Co-author (Student Research Assistant, TrustHLT Lab, TU Darmstadt, May 2023 – November 2024)