Gradient Immunity: Null-Space Resistance to Malicious Fine-Tuning
Yuxuan Huang, Xingyu Zeng, Tianhang Zheng, Chaochao Lu
cs.CR, cs.AI, cs.CL
Submitted: 2026-08-05
Code: https://github.com/OpenCausaLab/Gradient-Immunity
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Terminology
Sources
- GPT-4 Technical Report
- Safety-Tuned LLaMAs: Lessons From Improving the Safety of Large Language Models that Follow Instructions
- Adaptive Defense against Harmful Fine-Tuning for Large Language Models via Bayesian Data Scheduler
- Booster: Tackling Harmful Fine-tuning for Large Language Models via Attenuating Harmful Perturbation
- OSNIP: Breaking the Privacy-Utility-Efficiency Trilemma in LLM Inference via Obfuscated Semantic Null Space
- Harmful Fine-tuning Attacks and Defenses for Large Language Models: A Survey
- Defending Against Unforeseen Failure Modes with Latent Adversarial Training
- MetaDefense: Defending Finetuning-based Jailbreak Attack Before and During Generation
- AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models
- EvoDefense: Co-Evolving Black-Box Defense with Large Language Models
- The Llama 3 Herd of Models
- DeepSeek-V3 Technical Report
- An Early Warning of Emerging Biosecurity Risks in Frontier LLMs
- Not All Turns Matter: Credit Assignment for Multi-Turn Jailbreaking
- Decoupled Weight Decay Regularization
- Crossing the Reward Bridge: Expanding RL with Verifiable Rewards Across Diverse Domains
- HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal
- Reinforcement Learning with Verifiable Rewards: GRPO's Effective Loss, Dynamics, and Success Amplification
- Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications
- Antibody: Strengthening Defense Against Harmful Fine-Tuning for Large Language Models via Attenuating Harmful Gradient Influence
Related papers
- SoK: AI-Augmented Binary Reversing
- Relaxed Sender Anonymity for CBDC Interbank Settlement: A Zero-Knowledge Approach on Permissioned EVM
- Calibration-Family Overfit: Why Trusted Sabotage Monitors Don't Transfer Across Lineages
- Estimating the Social Cost of Corporate Data Breaches
- Efficient Fuzzy PSI under One-Sided Assumptions
- Sealing the Audit-Runtime Gap for LLM Skills