yw3408 / backdoor_detect

BBD:Bimodel Backdoor Defence

BBD is a Backdoor Defence Technique design for correctly recognize clean data and identify poisoned data as N+1 Label. To see the implementation, please open src/BBD.ipythnb, and run file

Group Member

Yuanrui Huang(yh2910)

Yihong Wang(yw3408)

Jiru Wang(jw6288)

Yichen Wang(yw4604)

Evironment

Python=>3.x

Jupyter Notebook

Tensorflow == 1.x(forneural_clease.ipythnb to run)

tensorflow-model-optimization

Keras => 2.x

h5py

HOWTO

To run source code in /src, please open codefiles in Jupyter Notebook and modify data and model path into your own customized path.

Fine_Pruning_Weight_Pruning.ipynb

• clean_data_filename => customized clean validation data path
• test_data_filename =>customized clean test data path
• poison_data_filename=>customized poison data path

Fine_Pruning_activation_based.ipynb

• clean_data_filename => customized clean validation data path
• test_data_filename =>customized clean test data path
• poison_data_filename=>customized poison data path

Entropy_filter.ipythnb

• clean_data_filename =>customized clean validation data path

• test_clean_data_filename =>customized clean test data path

• poison_data_filename => customized poison data path

neural_clease.ipythnb

• DATA_DIR => customized data folder

• DATA_FILE => customized clean data path

• MODEL_DIR => customized model folder

• MODEL_FILE => customized Badnet path

BBD.ipythnb

• DATA_DIR => customized data folder
• CLEAN_DATA_FILE => customized clean validation data path
• POISON_DATA_FILE=> customized poison data path
• MODEL_DIR => customized model folder
• BAD_MODEL_FILE => customized Badnet path
• CLEAN_MODEL_FILE => customized clean model path

All files can run idenpendently and with differnt purpose.

File Describtion

File	Input	Output	Describtion
neural_clease.ipythnb	<Badnet, Clean_Data>	potential_trigger
Fine_Pruning_Weight_Pruning.ipynb	<Badnet, Clean_Data>	Repaired_Model
Fine_Pruning_activation_based.ipynb	<Badnet, Clean_Data>	Repaired_Model
Entropy_filter.ipythnb	<Pruned Badnet, Clean_Data, Poison_Data>	Entropy Filter
BBD.ipythnb	<Badnet, Repaired Model>	Expected Classification

Reference

• [1] Bryant Chen, Wilka Carvalho, Nathalie Baracaldo, Heiko Ludwig, Benjamin Edwards, Taesung Lee, Ian Molloy, and Biplav Srivastava. 2018. Detecting backdoor attacks on deep neural networks by activation clustering.
• [2] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. 2019. Strip: A defence against trojan attacks on deep neural networks.
• [3] Yiming Li, Baoyuan Wu, Yong Jiang, Zhifeng Li, and Shu-Tao Xia. 2020. Backdoor Learning: A Survey.
• [4] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. 2018. FinePruning: Defending Against Backdooring Attacks on Deep Neural Networks.
• [5] B. Wang, Y. Yao, S. Shan, H. Li, B. Viswanath, H. Zheng, and B. Y. Zhao. 2019. Neural Cleanse: Identifying and Mitigating Backdoor Attacks in Neural Networks.
• [6] Jiale Zhang, Junjun Chen, Di Wu, Bing Chen, and Shui Yu. 2019. Poisoning attack in federated learning using generative adversarial nets.