Learn to Defend: Adversarial Multi-Distillation for Automatic Modulation Recognition Models

by [Zhuangzhi Chen], [Zhangwei Wang], [Dongwei Xu], [Jiawei Zhu], [Weiguo Shen], [Shilian Zheng], [Qi Xuan], [Xiaoniu Yang]

Main components

File name	Explanation
AMD.py	Execute the main code, with distillation model training.
CleanDatasets/CandidatesSelection.py	Code for extracting samples with correct model classification .
Attacks/*	Code for the adversarial attacks we use.
Defenses/*	Code for generating defense model.
model/*	Code for all model architecture we use in experiments.
raw_model_training/*	Code for model architecture and baseline training scripts.
Utils/dataset.py	Code for dataset preparation.
order/*	Code for running script .
args.py	Code for configuration.

Running Codes

How to train baseline models

python train.py --model CNN1D --dataset 128 --num_workers 8 --epochs 50

Alternatively, scripts can be used to batch train the model

bash raw_model_training/train.sh

How to train adversarial distillation models

To run the experiments, enter the following command.

python AMD.py \
         --save_root "model save root" \
         --t1_model "clean teacher model root" \
         --t2_model "adversarial teacher model root" \
         --s_init "student model root" \
         --dataset 128   --t1_name MCLDNN --t2_name Vgg16  --s_name mobilenet \
         --kd_mode logits --lambda_kd1 1 --lambda_kd2 1 --lambda_kd3 1 \
         --note \
         --gpu_index  

How to generate defense model

Defenses	Commands with default parameters
PAT	python PAT_Test.py --dataset=128--eps=0.06 --step_num=5 --step_size=0.03 --model CNN1D
TRADES	python train_trades.py --dataset=128--eps=0.06 --step_num=5 --step_size=0.03 --model CNN1D
DD	python DD_Test.py --dataset=128 --temp 30.0 --model CNN1D

How to evaluate model with attack method

1. Obtain samples with correct model classifacation. Clean samples will saved at /CleanDatasets

cd ./CleanDatasets
python CandidatesSelection.py --dataset=128 --number=$1  --gpu_index $2 --model $3  --location $4

2. Use the attack methods provided below. Adversarial samples will saved at /AdversarialExampleDatasets

Attacks	Commands with default parameters
PGD	PGD_Generation.py --dataset=128 --epsilon=0.15 --epsilon_iter=0.03 --num_steps 15 --model=CNN1D
UMIFGSM	python UMIFGSM_Generation.py --dataset=128 --epsilon=0.15 --model=CNN1D
AA	python AutoAttack.py --dataset=128 --epsilon=0.15 --model=CNN1D
DF	python DeepFool_Generation.py --dataset=128 --max_iters=15 -- overshoot=0.02 --model=CNN1D
SQUARE	python square.py --Linf=0.3 --num_queries=3000 --model CNN1D

Package Requirements

To ensure success running of the program, the versions Python packages we used are listed in requirements.txt.To align the versions of your packages to this file, simply run:

pip install -r requirements.txt