AI Paper Atlas

Research map / Federated learning and privacy / Privacy attacks and secure learning

OSNIP: Balancing the Privacy-Utility-Efficiency Trilemma in LLM Inference via Obfuscated Semantic Null Space

Zhiyuan Cao, Zeyu Ma, Chenhao Yang, HAN ZHENG, Mingang Chen

ICML 2026 · Federated learning and privacy › Privacy attacks and secure learning · approach: Encryption-based private learning

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 changedWhereMeasuredEffect
add the key-conditioned perturbation (ablation)OSNIP · Adaptive random-key attack on Qwen3-32BASR: 100.00% -> 0.00%Improved
add the diversity loss L_div (ablation)OSNIP · Adaptive random-key attack on Qwen3-32BASR: 100.00% -> 0.00%Improved
add oracle key knowledge for the attacker (ablation)OSNIP · Adaptive oracle-key attack on Qwen3-32BASR: 0.00% -> 100.00%Hurt
use OSNIP instead of Capeprivacy-preserving inference · KNN-Top10 attack on Qwen3-32BASR: 0.578 -> 0.000Improved
use OSNIP instead of DYNTextprivacy-preserving inference · 10-task average on Qwen3-32Baccuracy (Avg): 0.428 -> 0.748Improved
use OSNIP instead of InferDPTprivacy-preserving inference · WikiText-2 on Qwen3-32Bperplexity: 1155.74 -> 6.94Improved
increase model scale from Llama-3.2-1B to Qwen3-32BOSNIP · KNN-Top10 attackASR: 0.066 -> 0.000Improved
increase model scale from Llama-3.2-1B to Qwen3-32BOSNIP · WikiText-2perplexity increase over non-private: +31% -> +3%Improved
use OSNIP instead of InferDPTsummarization · CNN/DailyMail ROUGE-L on Qwen3-32BROUGE-L: 0.197 -> 0.182Hurt
use OSNIP instead of InferDPTsummarization · CNN/DailyMail on Llama-3.2-1BBERTScore: 0.868 -> 0.840Hurt
exclude stop words from vocabulary-matching attack ASROSNIP · Vocabulary-matching attack, Qwen3-32B L10ASR: 18.76 -> 7.23Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy