AI Paper Atlas

Research map / Large language models / Reasoning and chain of thought

SLIM: Secure and Efficient Inference for Large Language Models on Untrusted Devices via TEEs

Wei Wang, Zihao Guan, Xing Zhou, Yan Ding, Yusong Tan, Jie Yu, Bao Li

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient LLM serving

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SLIM, a scheme for secure and efficient LLM inference on untrusted devices that uses TEEs together with weight obfuscation to prevent weight extraction.

Problem

Deploying LLMs on untrusted hardware risks weight extraction and unauthorized replication; existing TEE-based weight obfuscation schemes struggle to provide both strong security guarantees and high performance.

Approach

Uses Trusted Execution Environments with obfuscated model weights to protect LLM weights during inference on untrusted devices, aiming to combine strong security with low overhead (details not available in the provided abstract).

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use SLIM instead of per-vector scaling/permutation obfuscation (SOTER, ShadowNet, TransLinkGuard, CoreGuard)TEE obfuscation schemes · GPT2-Small MNLI after statistical attackattack-recovered accuracy (%): 80.94 -> 40.65 (CoreGuard vs SLIM)Improved
increase Householder block size k beyond 32 (ablation)SLIM · Qwen3 GoEmotions, LLaMA3.2 FinQAattack-recovered accuracyHurt
use block sizes 2, 8 or 32 (ablation)SLIM · Qwen3 GoEmotions, LLaMA3.2 FinQAdefense effectNo clear effect
increase number of transformation rounds (ablation)SLIM · Qwen3 GoEmotions, block size 32 and 64model securityImproved
increase block size k (throughput vs KL divergence) (ablation)SLIM · Throughput with 128-token inputsthroughput / KL divergenceMixed
propagate authorization through layers (one TEE interaction) instead of per-layer TEE interactionSLIM vs ArrowCloak/GroupCover · TTFT, four causal LMsTTFT: 4.49x-13.80x (ArrowCloak), 5.68x-21.31x (GroupCover) overhead relative to SLIMImproved
use SLIM with orthogonal mixing in FP16SLIM vs ArrowCloak/GroupCover · Top-1 token accuracy, FP16Top-1 token accuracy: ArrowCloak 0% -> SLIM 95.2% on Gemma3Improved
increase model size from GPT2-Small to Qwen3-4BSLIM · TTFT ratio to Pure-GPUTTFT ratio: 3.66x -> 1.38xImproved
give attacker up to 60% of data for Full-Tuning or LoRA fine-tuning (ablation)SLIM · Qwen3 GoEmotions, LLaMA3.2 FinQA, GPT2 MNLI, Gemma3 WICattack-recovered accuracyNo clear effect

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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