Model Immunization Beyond Condition Numbers: The Importance of Plateau Regions
- Published
- Record updated
Summary
Researchers propose a model immunization method that defends against harmful fine-tuning by extending prior curvature-based work to also prioritize plateau regions. Their algorithm relies on efficient Hessian estimation and bounds on the extremal singular values of the Hessian, avoiding the high cost of earlier approaches. Experiments across fine-tuning-as-a-service, open-weight, and robust unlearning settings report robust defense against adversarial fine-tuning while largely preserving general utility and downstream trainability.
Mitigation
The proposed immunization framework itself is the defense: an immunization algorithm based on efficient Hessian estimation and bounds on the extremal singular values of the Hessian. Code is available at https://github.com/minhquoc0712/BECON.
Related items
- MediumGemini AI agents repeatedly bypassed safeguards to attack real websiteSimilar attack · BleepingComputer
- CriticalCVE-2026-120699: CubeSandbox host path bind mount allows host file access via POST /sandboxesSimilar attack · NVD/CVE Database
- MediumGoogle warns that fully-automated AI attacks are coming soonSimilar attack · BleepingComputer
- InfoBeyond the Patch: Reconstructing an Incomplete-Fix Patch Bypass in LLM Agent Frameworks with Control-Gated, Benign ProbesSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
- InfoCentered Ablation: Why Last-Layer Directional Ablation Under-Fires or Collapses, and a Closed-Form FixSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)