Model Immunization from a Condition Number Perspective

2025/05/29 by Amber Yijia Zheng, Site Bai, Zheng, Amber Yijia +5 · 2 citations
Biochemistry, Genetics and Molecular Biology · Medicine · Social Sciences · #FOS: Computer and information sciences #Influenza Virus Research Studies #Machine Learning (cs.LG) #Vaccine Coverage and Hesitancy #vaccines and immunoinformatics approaches

paper · pdf · doi:10.48550/arxiv.2505.23760

openalex publication_date 2025/05/29 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

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.

Citations

• Unlearning Isn't Deletion: Investigating Reversibility of Machine Unlearning in LLMs
• Antidistillation Sampling
• Multi-concept Model Immunization through Differentiable Model Merging
• Probe-Me-Not: Protecting Pre-trained Encoders from Malicious Probing
• (Almost) Smooth Sailing: Towards Numerical Stability of Neural Networks Through Differentiable Regularization of the Condition Number
• Learning to Obstruct Few-Shot Image Classification over Restricted Classes
• Harmful Fine-tuning Attacks and Defenses for Large Language Models: A Survey
• Weight Conditioning for Smooth Optimization of Neural Networks
• Generative AI Misuse: A Taxonomy of Tactics and Insights from Real-World Data
• Reconstruction Attacks on Machine Unlearning: Simple Models are Vulnerable
• IMMA: Immunizing text-to-image Models against Malicious Adaptation
• Removing RLHF Protections in GPT-4 via Fine-Tuning
• Structured Semidefinite Programming for Recovering Structured Preconditioners
• Unified Concept Editing in Diffusion Models
• Forget-Me-Not: Learning to Forget in Text-to-Image Diffusion Models
• Ablating Concepts in Text-to-Image Diffusion Models
• Erasing Concepts from Diffusion Models
• Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models
• PUMA: Performance Unchanged Model Augmentation for Training Data Removal
• Remember What You Want to Forget: Algorithms for Machine Unlearning
• Learning Transferable Visual Models From Natural Language Supervision
• Variational Bayesian Unlearning
• An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
• Fast and Near-Optimal Diagonal Preconditioning
• Descent-to-Delete: Gradient-Based Methods for Machine Unlearning
• PyTorch: An Imperative Style, High-Performance Deep Learning Library
• Certified Data Removal from Machine Learning Models
• A Comprehensive Survey on Transfer Learning
• A Comprehensive Survey on Transfer Learning
• The Malicious Use of Artificial Intelligence: Forecasting, Prevention, and Mitigation
• Sticking the Landing: Simple, Lower-Variance Gradient Estimators for\n Variational Inference
• Deep Residual Learning for Image Recognition
• Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation
• Выпуклая оптимизация

Cited by

• Knowledge Distillation Detection for Open-weights Models
• Gradient Surgery for Safe LLM Fine-Tuning

Related

• Multi-concept Model Immunization through Differentiable Model Merging
• Modelling Immunity in Agent-based Models
• Effective vaccination strategy using graph neural network ansatz
• Physics-Informed Neural Networks for Optimal Vaccination Plan in SIR Epidemic Models
• Overcoming Non-Submodularity: Towards Constant Approximation for Network Immunization