Reproduce “Inspecting the Representation Manifold of Differentially-Private Text” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Stefan Arnold
Researchers from Friedrich-Alexander-Universität Erlangen-Nürnberg introduce a geometric analysis framework for evaluating differentially-private text paraphrasing methods by measuring intrinsic dimensionality shifts in text representation manifolds, revealing that masked language model approaches better preserve structural complexity compared to word-level or causal methods across multiple privacy budgets.
Key insights
- Word-level DP methods cause more severe distortions to representation geometry than sentence-level approaches
- Sequential error propagation in causal language models leads to increased geometric distortion
- Masked language model approaches better preserve text structure while maintaining privacy guarantees
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.