The repository contains the code for ACL 2026 (main) paper "SharedRequest: Privacy-Preserving Model-Agnostic Inference for Large Language Models".

1. Installation

# Install dependencies
pip install -r requirements.txt

Then cd to the src folder.

2. Offline Training

Train a local discrimination model before online inference

./scripts/run_train_discrim.sh

Can edit the --device or --model_name based on the configuration of gpu or model path.

After training, a discrimination model will be stored in the ./model path.

3. Online Inference

3.1 Evaluate privacy protection (ASR & F1)

Step 1: Generate fake queries

First we generate a file containing the sample fake queries:

python -m online_infer.filter --model_name FacebookAI/roberta-base --device cuda:0

A file named "fake_attr_FacebookAI-roberta-base_1.json" will be created in the result/mmlu_fina folder.

Step 2: Evaluate attack accuracies for confusion strategy

We evaluate the attack success rate and F1 for our strategy:

./scripts/run_attack_confuse.sh

The script conducts inference attack using meta-llama/Llama-3.2-1B and FacebookAI/roberta-large as the attack models. The results will be stored in "result/logs/id-attack-attackmodel-{attack model}-discrimmodel-roberta-base-mmlu_fina-1.log" files.

Step 3: Evaluate attack accuracies for baseline method

For baseline method, we do not utilize the discrimination model to choose the fake queries. Instead, we randomly sample a set of queries from the attribute database.

The attack success rate and F1 is obtained from:

./scripts/run_attack_base.sh

The script conducts inference attack using meta-llama/Llama-3.2-1B and FacebookAI/roberta-large as the attack models. The results will be stored in "result/logs/id-attack-base-{attack model}-mmlu_fina.log" files.

3.2 Evaluate time and throughput

Step 1: Query time & troughput for our method:

./scripts/run_infer_confuse.sh

The script will output query time and throughput for gpt-4.1-mini and gpt-3.5-turbo models, with time results stored in the "./result/logs/confuse-query-{query_model}-mmlu_fina-1-roberta-base.log" file.

For adaptation to changan setting, can change the --query_online_model to "changan", and modify the functions (create_client & get_response) in ./utilts/query_utils.py.

Step 2: Query time & troughput for non-private setting:

./scripts/run_infer_base.sh

The script will output query time and throughput for gpt-4.1-mini and gpt-3.5-turbo models, with time results stored in the "./result/logs/base-query-mmlu_fina-{query_model}.log" file.

For adaptation to changan setting, can change the --query_online_model to "changan".