AIPDC：Artificial Intelligence Poisoned Data Clean System

Dependencies

This code sample is written in tensorflow and tf.keras. The code has been tested on tensorflow==1.14 and keras==2.2.5.

System Module

Getting Started:

1. clone the repo:

   git clone https://github.com/RidiculousDoge/AIPDC.git

2. Download sample dataset(GTSRB Dataset):

   cd AIPDC
   ./download_data.sh

3. Run badnet trainer and insert forward triggers into the model

   mkdir output
   python gen_backward/train_badnet.py --train --poison-type FF --poison-loc TL --poison-size 8 --epochs 10 --display 

   Currently supported options:

   • train: marks whether to train the model
   • poison-type: forward trigger type. Currently support FF & whitesquare
   • poison-loc: forward trigger location. Currently support TL(Top Left) & BR(Bottom Right)
   • poison-size: forward trigger size
   • epochs: train epochs
   • display: to show train plot or not

   to evaluate the forward-trigger insert process, run

   python gen_backward/eval_forward.py --checkpoint [your model]

   to check the insert influence.

4. Run backward trigger generator and generate mask & pattern data

   mkdir backward_triggers
   python3 gen_backward/snooper.py --checkpoint output/badnet-FF-TL-8-10-0.97.hdf5

   • checkpoint: the model saved in Step3.

   After the implementation, mask_FF_TL_8.npy & pattern_FF_TL_8.npy will be saved to the directory/backward_triggers

5. Apply Poisoned Data Cleanse Algorithm

   mkdir retrain_models
   python clean_and_retrain/data_clean.py --checkpoint output/badnet-FF-TL-8-10-0.97.hdf5 [optional:--narrow,--retrain]

   Currently support parameters:

   • narrow: whether the mask and pattern was trained with narrowed dataset.
   • retrain: after detecting data, whether to retrain the model with the eliminated dataset.

6. Evaluate the retrained model

python3 clean_and_retrain/eval_clean.py --checkpoint output/retrain-FF-TL-8-06-0.98.hdf5

7. Sample result

   1. Evaluate the effect of data cleaner:

      

   2. Evaluate the effect of retrainer: