Drone Authentication via Acoustic Fingerprint

• Basic tutorial about this project.
• Platform specification.
• Package requirements.
• Link of drone audio dataset.

The public GitHub link will be released in the final submission of our paper. For the artifacts evaluation, we provide this GitFront link.

Dataset

Drone_Audio (temporary).

DOI of the drone audio recordings. This link will be activated in the final submission. The content is the same as the temporary link above.

Requirements

The platform is Window 10. (Ubuntu 20.04 may be fine.)

Please install Anaconda at first. If you do not have Anaconda, please see the following link.

Then create a new environment: conda env create -f environment.yml

You need about 80G of storage space to generate the PKL dataset and models.

Audio Dataset Explanation

• DS1: this dataset contained drone audio from No. 1 - No. 8.
• DS2: this dataset contained drone audio from No. 1 - No. 24.
• DS1N: we added AWGN to DS1 with 0 dB SNR to create DS1N. The lengths of the corresponding drone audio in DS1N and DS1 are equal to each other.
• DS2N: we added AWGN to DS2 with 93 levels of SNR ranging from -8.00 dB to 15.00 dB with a step of 0.25 dB to create DS2N. The size of DS2N is 93 times larger than the size of DS2. In other words, each level of SNR creates a new subset in DS2N, whose size is equal to DS2.

Folder Description

dataset_build

Build MFCC dataset based on collected drone audio. The format of the generated dataset is ".pkl".

experiment

The code of all experiments, which are mentioned in the paper.

runners

These files help with the training and evaluation process.

toolbox

Some modules can be used in different programs.

Explaination of Config file

• originData_path: The root of all drone audio.
• output_path: Where to output/obtain the model.
• csv_savePath: Where to save the evaluation results.
• pkl_savePath: Where to save the dataset of extracted MFCC features.

At the Beginning

Download the drone audio dataset, then change all originData_path in all config files to the root of download drone audio.

Experiment of Frame Length Change

1. Run ./dataset_build/pkl_gen_timeVar.py to generate the dataset in .pkl format. This dataset is created from DS1.
   • Default dataset storage path is ./pkl_dataset/1_timeVar.
2. Run ./experiment/timeVar/train_all_model.py to train 8 different ML models on all generated datasets.
   • Default model storage path is ./trained_model/1_timeVar.
3. Run ./experiment/timeVar/eval_all_model.py to obtain the accuracy of each model on test set.
   • Default csv storage path is ./result/1_timeVar.

Filter-varying Experiment

Overview

There are 9 config files for training different models.

• config_filterVar_1_oneThird
• config_filterVar_1_twoThirds
• config_filterVar_1_all
• config_filterVar_2_oneThird
• config_filterVar_2_twoThirds
• config_filterVar_2_all
• config_filterVar_3_oneThird
• config_filterVar_3_twoThirds
• config_filterVar_3_all

PKL Generation

1. Run ./dataset_build/pkl_gen_filterVar.py to generate the dataset in .pkl format. This dataset is created from DS1.

Training and Evaluation

1. Change the name of config file in line 90 of ./experiment/filterVar/train_all_model_filterVar.py to train different models.
2. Change the name of config file in line 104 of ./experiment/filterVar/eval_all_model_filterVar.py to evaluate different models.
3. Run ./experiment/filterVar/train_all_model_filterVar.py.
   • Default model storage path is ./trained_model/2_filterVar/8d_x_xxxx.
4. Run ./experiment/filterVar/eval_all_model_filterVar.py.
   • Default csv storage path is ./result/2_filterVar/8d_x_xxxx.

Filter-varying Experiment with AWGN

PKL Generation

1. Run ./dataset_build/pkl_gen_filterVar_noise.py to generate the dataset in .pkl format. This dataset is created from DS1N.

Evaluation

1. Change the name of config file in line 107 of ./experiment/filterVar/eval_all_model_filterVar_noise.py to evaluate different models.
2. Run ./experiment/filterVar/eval_all_model_filterVar_noise.py.
   • Default csv storage path is ./result/3_filterVar_noise/8d_x_xxxx.

Authentication of 24 drones without AWGN

This experiment are conducted on DS2.

Training and Evaluation

1. Run ./experiment/noiseVar/train_all_model_noNoise.py.
   • Default model storage path is ./trained_model/4_noiseVar/noNoise.
2. Run ./experiment/noiseVar/eval_all_model_noNoise.py.
   • The results will be shown in the console.

Authentication of 24 drones with AWGN

This experiment are conducted on DS2N.

PKL Generation

1. Run ./dataset_build/pkl_gen_noiseVar.py to generate the dataset in .pkl format. This dataset is created from DS2N.

Evaluation

1. Run ./experiment/noiseVar/eval_all_model_noiseVar.py.
   • Default csv storage path is ./result/4_noiseVar.

Security Study (Attack)

1. Run ./dataset_build/pkl_gen_base.py to generate the dataset in .pkl format. This dataset is created from DS2.
2. Run ./experiment/attack/train_attack.py.
3. Record dic_reg, dic_attack, dic_bg shown in console.
   • For example
4. Change the value of args.bg_type and args.attack_type in ./experiment/attack/evaluate_attack.py to dic_bg and dic_attack, respectively.
   • For example
5. Run ./experiment/attack/evaluate_attack.py.
   • The results will be shown in console.