Research map / Federated learning and privacy / Differential privacy
PrivCode++: Latent-Conditioned Differentially Private Code Generation for Comprehensive Guarantees
Zheng Liu, Chen GONG, Terry Yue Zhuo, Zhou Yang, Kecen Li, Wenlong Meng, Xinwen Hou, Yu Liu, Xiaochen Li
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 latent conditioning (PrivLC) vs unconditional generation (ablation) | PrivCode++ · Instruct-HumanEval (avg of 4 models) | average pass@1: 31.1 -> 52.0 | Improved |
| add latent alignment between code and prompt latents (ablation) | PrivCode++ · Complete-MBPP (avg of 4 models) | average pass@1: 47.7 -> 68.7 | Improved |
| raise latent dimension from 256 to 768 (ablation) | PrivCode++ · Instruct HumanEval (avg of 4 models) | average pass@1: 43.4 -> 52.0 | Improved |
| raise latent dimension from 768 to 2048 (ablation) | PrivCode++ · Instruct HumanEval (avg of 4 models) | average pass@1: 52.0 -> 47.9 | Hurt |
| use latent-conditioned synthesis instead of prompt-conditioned synthesis with public prompts | PrivCode++ vs PrivCode · BigCodeBench Complete Full (Qwen2.5-Coder-7B) | pass@1: 27.9 -> 43.3 | Improved |
| use latent-conditioned synthesis instead of PrivCode relaxed-privacy | PrivCode++ vs PrivCode · HumanEval Instruct (DS-Coder-6.7B) | pass@1: 56.1 -> 41.5 | Hurt |
| use PrivCode++ instead of DP instruction fine-tuning (DPFT) | PrivCode++ vs DPFT · HumanEval Instruct (Qwen2.5-Coder-7B) | pass@1: 62.2 -> 68.3 | Improved |
| condition on prompt embeddings alone instead of code latents (ablation) | PC-PromptEmb vs PrivCode++ · HumanEval Instruct (Qwen2.5-Coder-7B) | pass@1: 68.3 -> 54.9 | Hurt |
| use PrivCode++ latent conditioning instead of PrivCode with sensitive prompts | PrivCode++ · Joint canary leakage, count=100, Qwen2.5-Coder-7B | leakage rate: 20% -> 0% | Improved |
| increase privacy budget epsilon (ablation) | PrivCode++ · HumanEval/MBPP average pass@1 | pass@1 | Improved |
| use public-pretrained prefix embedding (PC-PreEmb) instead of DP-trained latent (ablation) | PC-PreEmb vs PrivCode++ · HumanEval Complete (Qwen2.5-Coder-7B) | pass@1: 64.0 -> 51.2 | Hurt |
Closest papers by meaning
- Instruction Tuning for Secure Code Generation (ICML 2024 · similarity 0.88)
- Privacy-Preserving Instructions for Aligning Large Language Models (ICML 2024 · similarity 0.88)
- DP-OPT: Make Large Language Model Your Privacy-Preserving Prompt Engineer (ICLR 2024 · similarity 0.87)
- Data-adaptive Differentially Private Prompt Synthesis for In-Context Learning (ICLR 2025 · similarity 0.87)
- Efficient and Privacy-Preserving Soft Prompt Transfer for LLMs (ICML 2025 · similarity 0.86)
- Scaling Laws for Differentially Private Language Models (ICML 2025 · similarity 0.85)
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.
