🧪 Test?View on arXiv
Curvature Cryptanalysis of Smooth Transformer Feed-Forward Networks
Not specified in the provided content
cryptanalysistransformermodel extractionfeed-forward networks
2608.28843
Builder Relevance
1h ago70%
Abstract
The paper demonstrates a method for extracting structural information from smooth two-layer feed-forward networks using second-order leakage channels.
Reality Card
Core Claim
The study establishes that only 16 projected Hessians can recover hidden FFN directions with high alignment, achieving over 93% top-1 agreement in functional extraction.
Method / Result
Achieved an average absolute cosine alignment above 0.94 with only 16 projected Hessians from 8193 black-box queries.
Limitations
Output rounding and Gaussian noise significantly reduce recovery effectiveness under fixed attack configurations.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.
No verified implementation link has been attached yet. AIBuzzHub will keep this panel separate from unverified search results.