Stories about Transformer
1 related stories
Curvature Cryptanalysis of Smooth Transformer Feed-Forward Networks
AI InsightThis study shows that smooth two-layer FFNs expose hidden structure through a second-order leakage channel using projected input Hessians, under chosen-input raw-output access, without parameters, gradients, or internal activations. Compared to prior extraction methods that required white-box or internal state access, this systematically exploits curvature information and establishes identifiability and stability conditions.Key TakeawayModel extraction drops from requiring internal access to black-box queries plus second-order curvature analysis.Why It MattersThis reveals a new structural leakage surface for smooth-activation FFNs in pure black-box settings, potentially threatening proprietary model IP and prompting defenses focused on curvature obfuscation.Who's Affected- AI ResearchersCurvature leakage offers a new analysis framework for model extraction and identifiability.
- Cybersecurity PractitionersNeed to assess practical threats of second-order Hessian-based attacks on deployed models.
- Model ProvidersCommercial APIs exposing raw outputs may face higher structural theft risk.
- EnterprisesNeed to reassess trust boundaries when consuming black-box FFN models from third parties.
What's NextWatch whether this leakage channel extends to deeper or normalized Transformers, and attack feasibility under realistic query budgets.Importance 75/100