Research map / Large language models / Instruction and fine-tuning
ShieldedCode: Learning Robust Representations for Virtual Machine Protected Code
Mingqiao Mo, Yunlong Tan, Hao Zhang, Heng Zhang, Yangfan He
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 |
|---|---|---|---|
| add contrastive training (FCL+PCL) to the DeepSeekCoder VMP model (ablation) | ShieldedCode · VMP corpus ablation L1 | Pass@1: 18.95 -> 32.17 | Improved |
| add polymorphic generation (hierarchical attention) on top of contrastive training (ablation) | ShieldedCode · VMP corpus ablation (avg) | Pass@1: 21.86% to 25.17% | Improved |
| add polymorphic generation on top of contrastive training (ablation) | ShieldedCode · VMP corpus ablation (avg) | Pass@10: 35.25% to 38.30% | Improved |
| add polymorphic generation, L4 (ablation) | ShieldedCode · VMP corpus ablation L4 | Pass@1: 14.80 -> 22.36 | Improved |
| apply ShieldedCode approaches instead of standard fine-tuning | Granite 3B Code 128K · HumanEval_compile (avg L1-L4) | Pass@1: 12.84 -> 17.91 | Improved |
| apply ShieldedCode approaches instead of standard fine-tuning | Granite 3B Code 128K · HumanEval_compile (avg L1-L4) | Pass@10: 19.41 -> 25.25 | Improved |
| ShieldedCode vs GPT-4o on L0 generation | ShieldedCode · HumanEval_compile L0 | Pass@1: 22.58 -> 26.95 | Improved |
| ShieldedCode vs jTrans (Linear Probe) on O0+L1 | ShieldedCode · BinaryCorp-VirtualAssembly O0+L1 | Recall@1: 0.333 -> 0.488 | Improved |
| ShieldedCode vs jTrans (Linear Probe) on O2+L1 | ShieldedCode · BinaryCorp-VirtualAssembly O2+L1 | Recall@1: 0.715 -> 0.309 | Hurt |
| ShieldedCode vs jTrans (Linear Probe) on O1+L1 | ShieldedCode · BinaryCorp-VirtualAssembly O1+L1 | Recall@1: 0.573 -> 0.306 | Hurt |
| ShieldedCode vs VMProtect 3.7 on manual reversal success | ShieldedCode · Manual reverse engineering study | success rate: 67% -> 17% | Improved |
| ShieldedCode vs VMProtect under angr symbolic execution | ShieldedCode · angr symbolic execution | completion rate: 31% -> 3% | Improved |
Closest papers by meaning
- SecCodePRM: A Process Reward Model for Code Security (ICML 2026 · similarity 0.87)
- ProSec: Fortifying Code LLMs with Proactive Security Alignment (ICML 2025 · similarity 0.86)
- CODE REPRESENTATION LEARNING AT SCALE (ICLR 2024 · similarity 0.85)
- Instruction Tuning for Secure Code Generation (ICML 2024 · similarity 0.84)
- PurpCode: Reasoning for Safer Code Generation (NeurIPS 2025 · similarity 0.84)
- CodeMimicry: Exploiting Safety Generalization Lag in Large Language Models via Structured Code Completion (NeurIPS 2026 · similarity 0.84)
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.
