SHIELD: Fast, Practical Defense and Vaccination for Deep Learning using JPEG Compression


This is the code repository for the accepted KDD 2018 Applied Data Science paper: SHIELD: Fast, Practical Defense and Vaccination for Deep Learning using JPEG Compression. Visit our research group homepage Polo Club of Data Science at Georgia Tech for more related research!

The code included here reproduces our techniques (e.g. SHIELD) presented in the paper, and also our experiment results reported, such as using various JPEG compression qualities to remove adversarial perturbation introduced by Carlini-Wagner-L2, DeepFool, I-FSGM, and FSGM.

SHIELD overview YouTube video

Research Abstract and Contributions

The rapidly growing body of research in adversarial machine learning has demonstrated that deep neural networks (DNNs) are highly vulnerable to adversarially generated images. This underscores the urgent need for practical defense that can be readily deployed to combat attacks in real-time. Observing that many attack strategies aim to perturb image pixels in ways that are visually imperceptible, we place JPEG compression at the core of our proposed Shield defense framework, utilizing its capability to effectively “compress away” such pixel manipulation. To immunize a DNN model from artifacts introduced by compression, Shield “vaccinates” a model by re-training it with compressed images, where different compression levels are applied to generate multiple vaccinated models that are ultimately used together in an ensemble defense. On top of that, Shield adds an additional layer of protection by employing randomization at test time that compresses different regions of an image using random compression levels, making it harder for an adversary to estimate the transformation performed. This novel combination of vaccination, ensembling, and randomization makes Shield a fortified multi-pronged protection. We conducted extensive, large-scale experiments using the ImageNet dataset, and show that our approaches eliminate up to 94% of black-box attacks and 98% of gray-box attacks delivered by the recent, strongest attacks, such as Carlini-Wagner’s L2 and DeepFool. Our approaches are fast and work without requiring knowledge about the model.

Installation and Setup

Install Dependencies

This repository requires Cleverhans and TensorFlow-Slim image classification model library, please see their respective pages for instructions for installation.

Note: When installing TF-slim image models library, find models/research/slim, and put directory slim under utils. So that your directory structure should contain utils/slim.

Config Home Directory

In, fill in the home directory of your choice.

HOME_DIR = '' # eg. '/home/yourusername/'

Example usage:

The script can be used to perform (specified using –perform attack|defend|evaluate)

  1. attack - Attacks the specified model with the specified method(s)
  2. defend - Defends the specified attack images with the specified defense
  3. evaluate - Evaluates the specified model with the specified defended version of images.
python --use_gpu 0 --debug false --perform attack --models resnet_50_v2 --attacks fgsm,df
python --use_gpu 0 --debug true --perform evaluate --models resnet_50_v2 --checkpoint_paths /home/.../model.ckpt --attacks fgsm --defenses jpeg --attack_ablations '{"fgsm": [{"ord": Infinity, "eps": 2}]}' --defense_ablations '{"jpeg": [{"quality": 60}]}'

Video Demo

YouTube video demo

Paper PDF on arXiv

Paper PDF on arXiv


SHIELD: Fast, Practical Defense and Vaccination for Deep Learning using JPEG Compression.
Nilaksh Das, Madhuri Shanbhogue, Shang-Tse Chen, Fred Hohman, Siwei Li, Li Chen, Michael E. Kounavis, Duen Horng Chau.
ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) 2018. London, UK. Aug 19-23, 2018.


   title={SHIELD: Fast, Practical Defense and Vaccination for Deep Learning using JPEG Compression}, 
   author={Das, Nilaksh and Shanbhogue, Madhuri and Chen, Shang-Tse and Hohman, Fred and Li, Siwei and Chen, Li and Kounavis, Michael E and Chau, Duen Horng}, 
   booktitle={Proceedings of the 24nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining}, 


Name Affiliation
Nilaksh Das Georgia Tech
Madhuri Shanbhogue Georgia Tech
Shang-Tse Chen Georgia Tech
Fred Hohman Georgia Tech
Siwei Li Georgia Tech
Li Chen Intel Corporation
Michael E. Kounavis Intel Corporation
Polo Chau Georgia Tech