Model Immunity Is More Than Condition Number: The Importance of Plateau Regions
Model Immunization Beyond Condition Numbers: The Importance of Plateau Regions
Translation in progress — untranslated text is omitted.
The paper proposes an improved model immunization method to address the risk that harmful fine-tuning rapidly breaks LLM safety alignment and that deployed models are susceptible to post-release modification. Prior work slows harmful optimization by inducing pathological curvature, but it is difficult to scale because Hessian estimation is computationally expensive; this paper extends the theoretical framework to prioritize plateau regions beyond curvature, and designs an immunization algorithm based on efficient Hessian estimation and bounds on extreme Hessian singular values. The method is validated in three scenarios: fine-tuning-as-a-service where benign user data may be mixed with poisoned harmful data, open weights where an attacker fine-tunes only with harmful data, and robust unlearning where the immunized model is more resistant to relearning attacks after unlearning. The authors state that extensive experiments show the method robustly defends against adversarial fine-tuning while largely preserving general capabilities and downstream trainability. Code at https://github.com/minhquoc0712/BECON。
- Aligned LLMs fail after release when fine-tuned on a small amount of harmful data. The authors consider the previous model immunization approach based on the Hessian condition number to have insufficient theoretical grounding and to be difficult to scale to LLMs.
- The authors first prove that the convergence rate depends on both the condition number and the norm of harmful gradients, then propose BECON: it uses perturbed harmful gradients to construct a plateau loss that suppresses harmful gradients, uses a Gauss-Newton Gram matrix proxy to amplify the condition number, and jointly optimizes with the alignment loss.
- In the three scenarios of FTaaS (three models including Llama-2-7B), open weights (pure harmful fine-tuning sweeping learning rate and epoch), and WMDP-Bio against re-learning, the authors report that BECON has the lowest HS in most settings and maintains benign-task accuracy; as the attack budget increases, the HS of all defenses rises.
- The authors point out: harmful distributions outside the proxy data and stronger attacks will weaken the protection, immunization has utility loss and a one-time training cost, and the theory is limited to deterministic full-batch gradient descent; the main open-weight results are single-seed, and the HS review did not include additional human evaluation.