Reproduce “Privacy from Symmetry: Orthogonally Equivariant Transformers for LLM Inference” 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
Alexander Yukhimchuk, Andrey Shulga, Mladen Kolar, Martin Takáč
ConjFormer rotates client hidden vectors with a secret orthogonal transformation and modifies the transformer so the server can compute in that space. The client retains embeddings and the language-model head. On GPT-2 Medium fine-tuned on PubMed, top-10 recovery was 1.0% for Gram-based alignment versus 39.3% nearest-neighbor recovery on the original model. This does not rule out stronger attacks or protect visible invariants.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.