CRISP – Pretrained Model
This repository provides a pretrained model for CRISP, our framework for information-theoretic privacy-preserving edge–cloud inference.
Overview
CRISP uses privacy noise as side information at the client to recover accurate predictions from a server response to a privatized query. The model combines a ResNet-18 server with a small client-side decoder.
In our MNIST experiments with the Mixup mechanism at K = 16, CRISP offloads more than 99% of the neural-network computation to the server and achieves 97.1% client accuracy, while the evaluated adversary achieves only 21.7% accuracy in predicting the client digit label from the query. Beyond this empirical attack evaluation, the Mixup mechanism provides provable information-theoretic output-privacy guarantees.
For full details about the methodology, training setup, and data preprocessing, please refer to the main project repository:
👉 https://github.com/hehteram/CRISP
Contents
- Pretrained model checkpoint:
models/model_mixup_16.pth - Model weights for the CRISP Mixup configuration with K = 16
Usage
from huggingface_hub import hf_hub_download
import torch
path = hf_hub_download(
repo_id="hehteram/CRISP",
filename="models/model_mixup_16.pth"
)
model.load_state_dict(
torch.load(path, map_location="cpu", weights_only=True),
strict=True
)
Notes
- This model is provided for research and reproducibility purposes only.
- The dataset is MNIST. Use the matching preprocessing, Mixup query construction, and client decoding procedure from the main project repository.
License
Please refer to the main GitHub repository for licensing and usage terms.