safety security
Three Gradient Channels Gave the Tokens Away
AEGIS masked attention, embedding and MLP leakage paths in federated language-model fine-tuning.
Summary
AEGIS masked attention, embedding and MLP leakage paths in federated language-model fine-tuning.
The paper identifies three structural signals that gradient-inversion attacks can use to recover private training text: attention-projection subspaces, sparse embedding rows and an MLP expansion signal. AEGIS freezes or perturbs those backward paths and uses the same masked gradient locally and at the server boundary. Across 11 models and six datasets, the authors report near-zero token recovery with utility preserved or improved; deployment claims still depend on the tested attacks and threat model.
Why it matters
AEGIS masked attention, embedding and MLP leakage paths in federated language-model fine-tuning.
Limits and context
No additional limitation was separately recorded.
Key claims
AEGIS masked attention, embedding and MLP leakage paths in federated language-model fine-tuning.
Evidence: source-2026-08-22-008
Sources
- arXiv preprint 2608.19534arXiv · primary research
Corrections
No corrections have been recorded for this story.