Reproduce “SECNEURON: Reliable and Flexible Abuse Control in Local LLMs via Hybrid Neuron Encryption” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhiqiang Wang, Haohua Du, Junyang Wang, Haifeng Sun, Kaiwen Guo, Haikuo Yu, Chao Liu, Xiang-Yang Li
SECNEURON, developed by researchers from the University of Science and Technology of China, Beihang University, and Ocean University of China, proposes a method to embed access control directly into Large Language Models by selectively encrypting task-specific neurons using a hybrid CP-ABE and AES framework. This approach enables intrinsic capability control for local LLM deployments, resulting in over 40% accuracy reduction for unauthorized users on restricted tasks while maintaining near-original performance for authorized users.
Key insights
- Different neurons within a Large Language Model contribute distinctly to various tasks, enabling the selective encryption and decryption of specific functionalities.
- Embedding access control directly into the neural network structure, rather than relying on application-layer filters, prevents unauthorized capability activation even with full system access.
- A hybrid cryptographic framework combining CP-ABE for flexible policy management and AES for efficient neuron-level operations allows for granular and dynamic permission control.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.