Model Immunization from a Condition Number Perspective

TL;DR

This paper introduces a condition number-based framework to analyze and improve model immunization, enabling models to resist harmful task fine-tuning while maintaining utility on benign tasks, demonstrated on linear and deep models.

Contribution

It provides the first theoretical framework for model immunization using condition numbers and designs an algorithm to control this property during pre-training.

Findings

Effective immunization of linear models.

Successful application to non-linear deep networks.

Algorithm reduces vulnerability to harmful fine-tuning.

Abstract

Model immunization aims to pre-train models that are difficult to fine-tune on harmful tasks while retaining their utility on other non-harmful tasks. Though prior work has shown empirical evidence for immunizing text-to-image models, the key understanding of when immunization is possible and a precise definition of an immunized model remain unclear. In this work, we propose a framework, based on the condition number of a Hessian matrix, to analyze model immunization for linear models. Building on this framework, we design an algorithm with regularization terms to control the resulting condition numbers after pre-training. Empirical results on linear models and non-linear deep-nets demonstrate the effectiveness of the proposed algorithm on model immunization. The code is available at https://github.com/amberyzheng/model-immunization-cond-num.