Research map / Adversarial robustness and security / Attacks on text and graphs
LLA: Enhancing Security and Privacy for Generative Models with Logic-Locked Accelerators
You Li, Guannan Zhao, Yuhao Ju, Yunqi He, Jie Gu, Hai Zhou
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LLA (outlier-targeted permutation locking) instead of HPNN negation locking | Minitron/Llama/NeMo LLMs · pre-attack MMLU, 256 to 8192 protected neurons | MMLU | Improved |
| select protected block by first emergent outliers (add step i) (ablation) | LLA · Mistral NeMo Minitron 8B, pre-attack | MMLU and perplexity | Improved |
| select protected neurons by outlier-tracing score (add step ii) (ablation) | LLA · Mistral NeMo Minitron 8B, pre-attack | MMLU and perplexity | Improved |
| add Hadamard rotation matrix to obfuscate keys (ablation) | LLA · Mistral NeMo Minitron 8B, pre-attack | MMLU and perplexity | Improved |
| use permutation-based locking instead of negation-based locking (ablation) | LLA · gradient-based OG attack | key recovery fidelity: adversary recovers 100% of key bits (negation) vs at most 86% (permutation, 256 neurons) | Improved |
| increase permutation group size from 4 to 16 (ablation) | LLA · OG gradient attack, 4 models averaged | robustness: key bits per neuron 1.5 -> 3.5 | Improved |
| increase number of protected neurons (key length) (ablation) | LLA · gradient-based OG attack | max key fidelity: 86% -> 38% | Improved |
| use oracle-less (OL) instead of oracle-guided (OG) attack (ablation) | attack on LLA · group size 16, gradient | attack effectiveness | Hurt |
| use genetic-based instead of gradient-based key attack (ablation) | attack on LLA · OG, group size 16 | attack effectiveness | Hurt |
| add LLA locking module with 7,168 key bits | AI accelerator · computational overhead | overhead: less than 0.1% for 7,168 key bits | No clear effect |
Closest papers by meaning
- Breaking Agent Backbones: Evaluating the Security of Backbone LLMs in AI Agents (ICLR 2026 · similarity 0.84)
- MAP: MAsk-Pruning for Source-Free Model Intellectual Property Protection (CVPR 2024 · similarity 0.83)
- CoreGuard: Safeguarding Foundational Capabilities of LLMs Against Model Stealing in Edge Deployment (NeurIPS 2025 · similarity 0.83)
- SEC-bench: Automated Benchmarking of LLM Agents on Real-World Software Security Tasks (NeurIPS 2025 · similarity 0.83)
- FloatDoor: Platform triggered Backdoors in LLMs (NeurIPS 2026 · similarity 0.83)
- Catastrophic Jailbreak of Open-source LLMs via Exploiting Generation (ICLR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
