Protective Perturbation - MMSys 2022

In this project, we develop a protective perturbation generator at the user end, which adds perturbations to the input images to prevent privacy leakage. In the meantime, the image recognition model still runs at the service provider end to recognize the protected image without the need of being re-trained. The detail of this project is described in our recently accepted MMSys'22 paper:

Mengmei Ye, Zhongze Tang, Huy Phan, Yi Xie, Bo Yuan, Sheng Wei. Visual Privacy Protection in Mobile Image Recognition Using Protective Perturbation. ACM Multimedia Systems Conference (MMSys), June 2022.

A draft version of the paper is under the paper folder, and the camera-ready version will be available in April 2022.

This repo contains both code and instructions for the following three components:

• Set up the end-to-end image recognition system (both client and server) with the protective perturbation generator deployed at the client using our pre-trained model.
• Train our protective perturbation generator, with different target models and auxiliary models.
• Reproduce the experimental results reported in the paper.

Repository Hierarchy

 |-----server                            // server-end code, model training and results reproduction
       |-----cifar10_models              // pre-trained image recognition models for CIFAR10
       |-----generator_models            // pre-trained protective perturbation generators 
       |-----recv_from_phone             // folder to store images sent by the client
       |-----mobile_server.py            // server program
       |-----train.py                    // main program to train a protective perturbation generator
       |-----mobile_result_file_gen.py   // reproduce accuracy and SSIM results
       |-----cpu_gpu_timing.py           // reproduce timing results
       |-----module.py                   // the wrapper model of the generator to be trained
       |-----...                         // some other supplementary util codes
 |-----client                            // client-end code
       |-----app                         
             |-----src                   // source code of the Android APP
             |-----build.gradle          // gradle config
       |-----build.gradle                // gradle config
       |-----...                         // some other config files
 |-----paper
       |-----Paper-Protective_Perturbation-MMSys22.pdf  // camera ready version of the paper
       |-----README.md

1. Environment Setup

In this project, we use the same workstation (i.e., the server) to run the server program, develop and train the generators, analyze and produce the results reported in the paper, and develop and compile the Android client APP.

Hardware Requirements

Server

A workstation or desktop computer with GPU is required (additional modification to the code is needed if without GPU). In our experiments, an Nvidia RTX A6000 GPU is used.

Client

Theoretically, the APP can run on any Android phone w/ ARMv8 CPU and Android 8.0 or higher. The client APP has been fully tested on a Pixel 3 and a OnePlus 5T. All the results were collected on the Pixel 3. At least 2.5 GB free space on the phone is required to run the end-to-end system, and 6.0 GB is required if you want to collect all the experimental results.

Software Requirements

Server - server program, model training, results reproduction

This project is developed and tested in the following system:

Ubuntu 20.04 LTS
CUDA 11.5
Python 3.6
PyTorch 1.7.1+cu110
PyTorch-Lightning 1.2.3

Please set up the development environment by following the instructions below:

1. (IMPORTANT) We recommend using PyCharm to open the server folder directly, and set up everything for the server-end in the integrated terminal of PyCharm or using GUI.

2. Update Ubuntu first.

   sudo apt update
   sudo apt upgrade

3. Install CUDA (a version >=11.0 && <12.0 is okay) following the official instructions.

4. (Optional) Install Anaconda to setup a virtual environment. This will help protect your dev environment from being a mess. Setup the Conda virtual environment w/ Python 3.6. We highly recommend using PyCharm GUI to do so directly. If you prefer command line, try the following command to create a virtual environment called ProtectivePerturbation_MMSys22.

   conda create -n ProtectivePerturbation_MMSys22 python=3.6
   conda activate ProtectivePerturbation_MMSys22

5. Install the required libraries (in the virtual environment if you have one).

   pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html
   pip install matplotlib pytorch-lightning==1.2.3 wandb xlsxwriter

6. Download original_target_model.zip from this link and extract all the .pt files to <ProjectFolder>/server/cifar10_models/state_dicts/.

7. Download original_model.zip from this link and extract all the .ckpt files to <ProjectFolder>/server/generator_models.

Server - client APP development (skip this section if you do not need to modify the client APP)

1. Use Android Studio to open the client folder.

2. Download the CIFAR10 dataset (cifar10_test5000.zip), protective perturbation generator models (mobile_model.zip), and target models (mobile_target_model.zip) from this link.

   Extract these three zip files to <ProjectFolder>/client/app/src/main/assets. After extraction, the assets folder looks like below:

   assets
   ├── cifar10
   │   ├── 0000-7.png
   │   ├── 0001-5.png
   │   ├── 0002-7.png
   │   ├── ...
   ├── model
   │   ├── densenet121-X-densenet121.ptl
   │   ├── densenet121-X-googlenet.ptl
   │   ├── ...
   ├── PUT_ASSETS_HERE
   └── target_model
       └── mobilenet_v2.ptl
       └── ...
   

3. Modify the code, then Android Studio will handle the rest of things for you.

2. End-to-end System Setup and Execution Using Our Pre-trained Model

Set Up the Server

As shown in Fig 5 of our paper, our end-to-end system consists of a mobile app running on an Android smartphone and a server running on a workstation.

Figure 5 of the paper: Mobile system implementation

To set up the server, please follow the steps below:

1. Find the IP of the server. Open a Terminal and use ip addr to check the IP address. The result is something like below, where enp9s0 is the name of the NIC, and 192.168.6.164 is the IP address. You need to use the same IP address in the client APP.

   1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN group default qlen 1000
       link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
       inet 127.0.0.1/8 scope host lo
          valid_lft forever preferred_lft forever
   2: enp9s0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc mq state UP group default qlen 1000
       link/ether xx:xx:xx:xx:xx:xx brd ff:ff:ff:ff:ff:ff
       inet 192.168.6.154/24 brd 192.168.6.255 scope global dynamic noprefixroute enp9s0
          valid_lft 79207sec preferred_lft 79207sec
   

2. In server/mobile_server.py, change the HOST variable (str) to the IP address you find above. Change the PORT if 8081 is not available.

3. In server/mobile_server.py, change the DATA_DIR variable (str) to <Project Folder Path>/server/recv_from_phone, a folder that has existed in the root of the project folder.

4. Run server/mobile_server.py.

   cd server  # always make the server folder as your work folder when you try to run code on server!
   python mobile_server.py

   The server starts up successfully if you see the messages below:

   Global seed set to 0
   Init server successfully.
   Waiting for connection...
   

   Below is an example when the server receives a protected image from the client successfully:

   New connection <socket.socket fd=21, family=AddressFamily.AF_INET, type=SocketKind.SOCK_STREAM, proto=0, laddr=('192.168.6.154', 8081), raddr=('192.168.6.218', 38272)> ('192.168.6.218', 38272)
   Receiving image w/ size 3049
   Received recv_from_phone/0.png successfully
   Predicted label 4
   

Set Up the Client

To set up the client APP, please follow the steps below:

1. Install the APP.

   We recommend using ADB to install the APP. You have to enable USB debugging on the phone first, and then connect it to your machine via a USB cable. After that, install ADB on the machine.

   sudo apt install adb  # install adb on Ubuntu
   adb devices  # list connected devices

   If the phone is connected correctly, you will see outputs like below.

   List of devices attached
   XXXXXXXXX	device
   

   Then download the client APP ProtectivePerturbationDemo.apk from this link and install it using ADB.

   adb install ProtectivePerturbationDemo.apk

   You always can simply copy & paste it to your phone, and install it manually. But ADB is preferred as it is helpful to obtain protected images from the phone.

   Internet connection permission is required for this APP.

2. Make sure the server is up, and you already have the server IP address and the port number. Connect the phone to the same network as the server does.

3. Replace 192.168.6.154:8081 showing on the APP with Your_Server_IP_Address:Port_Number. The IP address and the port number of your server is set in server/mobile_server.py.

4. Click the button SINGLE DEMO. The client APP will pick one of the 5000 images from the dataset randomly and send it to the server. The original image and the protected image (generated by the APP) will be shown on the screen, and the protected one will be sent to the server. If everything is okay, the server will call the target model and return the result to the phone.

   An example is shown below. The first line Predicted label from server: 9 - truck is based on the returned result from the server. The second line shows some debug info. 2230-9.png is the name of the randomly-chosen figure, where 2230 is the index and 9 is the ground truth label. Label predicted on client 9 is the label classified by the target model running on the phone. This target model is the same as the one running on the server, and just for debug only.

   

   Example of the client demo

This demo uses Mobilenet_v2 as the target model. If you want to try another generator, simply change the number 25 in Line 75 and 76 to some other number i (0<=i<64) in MainActivity.java. Also, change target_model accordingly in the server/mobile_server.py of the server. Refer to Server - client APP development section for more instructions.

3. Model Training

Logger Setup

To set up wandb, please use the following command (See official document for more details):

wandb login

You can change logger to tensorboard, which can be used directly without an account.

Training Settings

In server/train.py, data_dir and gpu_id needs to be customized based on the setting of your environment. The target model is set by model_A, and the auxiliary model is set by model_B. All available models can be found in all_classifiers, which is defined in server/utils/all_classifiers.py. Change other parameters in MODULE args or TRAINER args if you'd like to customize the model.

Start Training

After you have modified the parameters in server/train.py, you can simply run it in PyCharm. You can also use command line instead to set the parameters and start the training, for example:

python train.py \
         --data_dir <your data dir> \
         --model_A resnet18 \
         --model_B resnet18 \
         --learning_rate 5e-3 \
         --gpu_id <your gpu id> \
         --description "resnet18-X-resnet18"

Note that for googlenet-X-vgg13_bn, the learning rate is 1e-3. For others, it is 5e-3. See Section 4.1.3 of the paper for details.

To retrain all the models, you can have the parameters set in server/train.py first, and then use all_in_one_script.sh to retrain all the models with only one line command.

bash all_in_one_script.sh <gpu id>

4. Reproducing the Experimental Results in the Paper

Effectiveness Evaluation

1. Generate all the protected images on the client APP first.

   WARNING: It takes about 2.5 hours on a Pixel 3. You should put the phone down and keep the APP in the foreground to avoid any interruption.

   All the 64 protective perturbation generators and 5000 test images have already been included in the prebuilt APK.

   To reproduce the results of our paper, you need to collect all the protected images of the 64 different perturbation generators. In total, there will be 320,000 images. To do so, just simply click GENERATE ALL PROTECTED IMG button, and wait until see Generating pics done. on the screen.

   To copy all the images from the phone to the workstation, use ADB:

   adb pull /sdcard/Android/data/rutgers.hwsel.ProtectivePerturbation/files <path to your data folder>

   WARNING: DO NOT USE MTP, as it is very unstable. If it interrupts frequently, try different USB ports or cables. It takes about an hour to pull all the 320,000 images from a Pixel 3 to the workstation.

2. Download the CIFAR10 dataset (cifar10_test5000.zip) from this link, and extract it to <data folder> (i.e., the folder where you adb pull all the images to in the previous step). Now in <data folder>, you should have 65 folders, under each there are 5000 images.

3. Modify data_dir in server/mobile_result_file_gen.py to the path of <data folder>.

4. Run server/mobile_result_collect.sh to calculate the results (accuracy for the target model & the auxiliary model, SSIM) for all the protected images.

cd server  # always in the server folder when you try to run something of the server
bash mobile_result_collect.sh

5. In the generated mobile_result_mmsys2022.txt, you can see the results are listed as follows:
   • Protective perturbation generator name (in the format of target model-X-auxiliary model.ptl)
   • Accuracy of the target model to classify protected images
   • Accuracy of the auxiliary model to classify protected images
   • SSIM between the original images and the protected images

Overhead Evaluation

1. In mobile_model.zip, you can check the size of all the protective perturbation generators.

2. For the timing results of CPU and GPU, you can use the code in cpu_gpu_timing.py to collect all the results. You have to change data_dir to an accessible folder where can put CIFAR-10 dataset, and the code will download it automatically.

3. For the timing results of mobile, follow the instructions below:

   WARNING: You should put the phone down and keep the APP in the foreground when measuring the time.

   Click the button MEASURE RUNNING TIME of the APP. After it is done, you will see Time measurement done. on the screen.

   To pull the results, use the following command:

   adb pull /sdcard/Android/data/rutgers.hwsel.ProtectivePerturbation/files/time_result.txt ./

   Check the time_result.txt you can see the generator name (the name before -X- is the target model) followed by the timing results (in milliseconds) for 5000 images.

   The timing results could be different from the paper, due to different phone models, APPs installed, and even current temperature. To get the most accurate result, you should modify the loop in the OnClickListener() of getTimeBtn (around line 175 in MainActivity.java) to run the time measurement one by one manually.

   To be more specific, after the measurement for modelList[i] (0<=i<8) is done, idle or even shut down the phone for five minutes, and then do modelList[i+8], until i>=64. We only do 8 models because there are 64 protective perturbation generators, but 8 target models in total, which means we just need to measure the time for 8 target models with the same auxiliary model.

5. Miscellaneous

Model Conversion - From Original to Mobile

We follow the official example to convert all the target models and protective perturbation generator models from original version to the mobile version, so that the converted version of the model can run on the smart phone with the help of PyTorch Mobile.

Check server/mobile_model_export.py for the detailed code. Remember to change GENERATOR_MODEL_SAVE_FOLDER and TARGET_MODEL_SAVE_FOLDER.

Uninstall the Client APP

Since the APP uses assets and internal usage to store the files, just simply uninstall the APP as normal, and everything will be gone.

6. Trouble Shooting

ERROR: Cannot uninstall 'certifi'.

When install matplotlib, you may see the error:

Attempting uninstall: certifi
Found existing installation: certifi 2016.9.26
ERROR: Cannot uninstall 'certifi'. It is a distutils installed project and thus we cannot accurately determine which files belong to it which would lead to only a partial uninstall.

Use the following commands to fix the error:

pip install certifi --ignore-installed
pip install matplotlib pytorch-lightning==1.2.3 wandb xlsxwriter

7. Cite Our Work

Mengmei Ye, Zhongze Tang, Huy Phan, Yi Xie, Bo Yuan, Sheng Wei. Visual Privacy Protection in Mobile Image Recognition Using Protective Perturbation. ACM Multimedia Systems Conference (MMSys), June 2022.

8. Contact

If you have any questions or any idea to discuss, you can email Zhongze Tang ([email protected]) directly.

9. References

In this project, the code and weights of classification models for CIFAR-10 dataset is from the project PyTorch models trained on CIFAR-10 dataset.

10. License

MIT