Reproduce “From Efficiency to Leakage -- Privacy Backdoor in Federated Language Model Fine-Tuning” 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

Shanghao Shi, Chaoyu Zhang, Heng Jin, Yang Xiao, Yevgeniy Vorobeychik, William Yeoh, Ning Zhang, Y. Thomas Hou and 1 more

NeuroImprint lets a malicious server recover client text by routing each sample’s update to an adapter neuron. It reverses final parameter changes. On GPT-2 SQuAD, SGD recovers 0.642 with similarity 0.988 ± 0.071, while AdamW recovers 0.734 with 0.682 ± 0.119. The attack requires server control and a crafted adapter; it does not break secure aggregation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.