Automated Side Channel Analysis of Media Software with Manifold Learning
Official implementation of USENIX Security 2022 paper: Automated Side Channel Analysis of Media Software with Manifold Learning.
Paper link: https://www.usenix.org/conference/usenixsecurity22/presentation/yuan
Extended version: https://arxiv.org/pdf/2112.04947.pdf
Note
Warning: This repo is provided as-is and is only for research purposes. Please use it only on test systems with no sensitive data. You are responsible for protecting yourself, your data, and others from potential risks caused by this repo.
Updates
- 2021 Oct 9. Released data, code and trained models.
Requirements
-
To build from source code, install following requirements.
pip install torch==1.4.0 pip install torchvision==0.5.0 pip install numpy==1.18.5 pip install pillow==7.2.0 pip install opencv-python pip install scipy==1.5.0 pip install matplotlib==3.2.2 pip install librosa==0.7.2 pip install progressbar
Then type
git clone https://github.com/Yuanyuan-Yuan/Manifold-SCA cd Manifold-SCA export MANIFOLD_SCA=$PWD
-
If you would like to build this repo from docker, see DOCKER and skip the following steps except for 6.1.
0. Output
We provide data and our trained models here.
Run the following scripts to produce outputs from these data samples
python output.py --dataset="CelebA"
python output_blind.py --dataset="CelebA"
python output_noise.py --dataset="CelebA"
You can choose dataset
from ["CelebA", "ChestX-ray", "SC09", "Sub-URMP", "COCO", "DailyDialog"]
. Results will be saved in output
.
We also provide all our outputs.
1. Datasets
We provide sampels of our processed data here.
CelebA
Download the CelebA dataset from here. We use the Align&Cropped Images
version.
After downloading the dataset, go to tool
. Then run
python crop_celeba.py --input_dir="/path/to/unzipped_images" --output_dir="/path/to/cropped_images"
to crop all images to size of 128*128
. We provide several examples in data/CelebA_crop128/image
.
ChestX-ray
Download the ChestX-ray dataset from here.
After downloading the dataset, go to tool
. Then run
python resize_chest.py --input_dir="/path/to/unzipped_images" --output_dir="/path/to/resized_images"
to convert all images to JPEG format and resize them to size of 128*128
. We provide several examples in data/ChestX-ray_jpg128/image
.
SC09 & Sub-URMP
Download the SC09 dataset from here and Sub-URMP dataset here.
We process audios in the Log-amplitude of Mel Spectrum (LMS) form, which is a 2D representation. Once the dataset is downloaded, go to tool
and run
python audio2lms.py --dataset="{SC09} or {Sub-URMP}" --input_dir="/path/to/audios" --output_dir="/path/to/lms"
to covert all audios to their LMS representations. Several examples are provided in data/SC09/lms
and data/Sub-URMP/lms
respectively.
COCO-Caption & DailyDialog
Download COCO captions from here. We use the 2014 Train/Val annotations
. After downloading you need to extract captions from captions_train2014.json
and captions_val2014.json
. We provide several examples in data/COCO/text/train.json
and data/COCO/text/val.json
.
Download DailyDialog dataset from here. After downloading you will have dialogues_train.txt
and dialogues_test.txt
. We suggest you store the sentences in json
files. Several examples are given in data/DailyDialog/text/train.json
and data/DailyDialog/text/test.json
.
Once the sentences are prepared, you need to build the corresponding vocabulary. Go to tool
and run
python build_vocab.py input_path="/path/to/sentences.json" --output_path="/path/to/vocabulary.json" --freq=minimal_word_frequency
to build the vocabulary. We provide our vocabularies in data/COCO/text/word_dict_freq5.json
and data/DailyDialog/text/word_dict_freq5.json
.
2. Target Software
Install libjpeg, hunspell and ffmpeg.
We already set up the three software in DOCKER.
3. Side Channel Attack
We analyze three common side channels, namely, cache bank, cache line and page tables.
3.1. Prepare Data
We use Intel Pin (Ver. 3.11) to collect the accessed memory addresses of the target software when processing media data. We already set up the Pin in DOCKER.
We provide our pintool in pin/pintool/mem_access.cpp
. Download Pin from here and unzip it to PIN_ROOT
(specify this path by yourself).
To prepare accessed memory addresses of libjpeg when processing CelebA images, first put pin/pintool/mem_access.cpp
into /PIN_ROOT/source/tools/ManualExamples/
and run
make obj-intel64/mem_access.so TARGET=intel64
to compile the pintool. Before collect the memory address, remember to run
setarch $(uname -m) -R /bin/bash
in your bash to disable ASLR. In fact, the ASLR does not affect our approach, so you can also drop the randomized bits of collected memory address.
Then put pin/prep_celeba.py
into /PIN_ROOT/source/tools/ManualExamples/
and set the following variables:
input_dir
- Directory of media data.npz_dir
- Directory where the accessed memory addresses will be saved. Addresses of each media data will be saved in.npz
format.raw_dir
- Directory where the raw output of our pintool will be saved. These files will be used for localize side channel vulnerabilities.libjpeg_path
- Path to the executable file of libjpeg.
You can speed up the progress by running multiple processes. Go to /PIN_ROOT/source/tools/ManualExamples/
and simply set variable total_num
in *.py
to the number of processes and run
python prep_celeba.py --ID=id_starting_from_1
to prepare data. Follow the same procedure for other datasets.
We provide our collected side channel records of all datasets here.
3.2. Map Memory Addresses to Side Channels
We map the collected memory addresses addr to side channels according to the following table.
CPU Cache Bank Index | CPU Cache Line Index | OS Page Table Index |
---|---|---|
addr >> 2 | addr >> 6 | addr >> 12 |
Set the following variables in tool/addr2side.py
.
input_dir
- Directory of collected.npz
files recording accessed memory addresses.cachebank_dir
- Directory of converted cache bank indexes.cacheline_dir
- Directory of converted cache line indexes.pagetable_dir
- Directory of converted page table indexes.
Then run
python addr2side.py
to get the side channels records. You can also speed up the progress by running multiple processes.
3.3. Reconstruct Private Media Data
You need to first customize following data directories in code/data_path.json
.
{
"dataset_name": {
"media": "/path/to/media_data/",
"cachebank": "/path/to/cache_bank/",
"cacheline": "/path/to/cache_line/",
"pagetable": "/path/to/page_table/",
"split": ["train", "test"]
},
}
To approximate the manifold of face photos from cache line indexes, go to code
and run
python recons_image.py --exp_name="CelebA_cacheline" --dataset="CelebA" --side="cacheline"
The recons_image.py
script approximates manifold using the train
split of CelebA dataset and ends within 24 hours on one Nvidia GeForce RTX 2080 GPU. Outputs (e.g., trained models, logs) will by default be saved in output/CelebA_cacheline
. You can customize the output directory by setting --output_root="/path/to/output/"
. The procedure is same for other media data (i.e., audio, text).
Once the desired manifold is constructed, run
python output.py --exp_name="CelebA_cacheline" --dataset="CelebA" --side="cacheline"
to reconstruct unknown face photos (i.e., the test
split). The reconstructed face photos will by default be saved in output/CelebA_cacheline/recons/
. This procedure is also same for audio and text data.
We use Face++ to assess the similarity of IDs between reconstructed and reference face photos. The online service is free at the time of writing, so you can register your own account. Then set the key
and secret
variables in code/face_similarity.py
and run
python face_similarity.py --recons_dir="../output/CelebA_cacheline/recons/" --target_dir="../output/CelebA_cacheline/target/" --output_path="../output/CelebA_cacheline/simillarity.txt"
The results will by default be saved in output/CelebA_cacheline/simillarity.txt
.
For ChestX-ray images, we use this tool to check the consistency between disease information of reconstructed reference images.
You can also evaluate the similarity between reconstructed and reference images by running tool/SSIM.py
.
python SSIM.py --K=1 --N=100 --recons_dir="../output/CelebA_cacheline/recons/" --target_dir="../output/CelebA_cacheline/target/" --output_path="../output/CelebA_cacheline/SSIM.txt"
The evaluation methods of audio data and text data are implemented in code/recons_audio.py
and code/recons_text.py
respectively. Note that the reconstructed audios are in the LMS representation, to get the raw audio (i.e., .wav
format), run
python lms2audio.py --input_dir="/path/to/lms" --output_dir="/path/to/wav"
If you want to use your customrized dataset, write your dataset class in code/data_loader.py
.
We also provide our trained models.
4. Program Point Localization
Once you successfully perform side channel attacks on the target softwares, you can localize the side channel vulnerabilities.
First customize the following variables in code/data_path.json
.
{
"dataset_name": {
"pin": "/path/to/pintool_output/",
},
}
Then go to code
and run.
python localize.py --exp_name="CelebA_cacheline" --dataset="CelebA" --side="cacheline"
The output .json
file will be saved in output/CelebA_cacheline/localize
. The results are organized as
{
"function_name; assmbly instruction; instruction address": "count",
}
The results of media software (e.g., libjpeg) processing different data (e.g., CelebA and ChestX-ray) are mostly consistent. We provide our localized vulnerabilities.
5. Perception Blinding
We provid the blinded media data here.
To blind media data, go to code
and run
python blind_add.py --meida="{image} or {audio} or {text}" --mask_weight=0.9 --mask="{mask_word} or {/path/to/mask_image_or_audio}" --input_dir="{/path/to/text.json} or {/foler/of/image_or_audio/}" --output_dir="{/path/to/text.json} or {/foler/of/image_or_audio/}"
To unblind media data, run
python blind_subtract.py --meida="{image} or {audio} or {text}" --mask_weight=0.9 --mask="{mask_word} or {/path/to/mask_image_or_audio}" --input_dir="{/path/to/text.json} or {/foler/of/image_or_audio/}" --output_dir="{/path/to/text.json} or {/foler/of/image_or_audio/}"
Run
python output_blind.py --dataset="CelebA"
# or
python output_blind.py --dataset="ChestX-ray"
to see reconstructed media data from side channels corresponding to blinded data.
6. Attack with Prime+Probe
We use Mastik (Ver. 0.02) to launch Prime+Probe on L1 cache of Intel Xeon CPU and AMD Ryzen CPU. We provide our scripts in prime_probe/Mastik
. After downloading Mastik, you can put our scripts in the demo
folder and run make
in the root folder to compile our scripts. We highly recommend you to set the cache miss threshold in these scripts according to your machines.
The Prime+Probe is launched in Linux OS. You need first to install taskset and cpuset.
6.1. Prepare Data
We assume victim and spy are on the same CPU core and no other process is runing on this CPU core. To attack libjpeg, you need to first customize the following variables in code/prime_probe/coord_image.py
pp_exe
- Path to the executable file of our prime+probe script.input_dir
- Directory of media data.side_dir
- Directory where the collected cache set accesses will be saved.libjpeg_path
- Path to the executable file of libjpeg.TRY_NUM
- Repeating times of processing one media using the target software.PAD_LEN
- The length that the collected trace will be padded to.
The script coord_image.py
is the coordinator which runs spy and victim on the same CPU core simultaneously and saves the collected cache set access.
Then run
sudo cset shield --cpu {cpu_id}
to isolate one CPU core. Once the CPU core is isolated, you can run
sudo cset shield --exec python run_image.py -- {cpu_id} {segment_id}
The script run_image.py
will run coord_image.py
using taskset. Note that we seperate the media data into several segments to speed up the side channel collection. The segment_id
starts from 0. The procedure is same for other media data.
We provide our logged side channels.
6.2. Reconstruct Private Media Data
First customize the following variables in code/data_path.json
.
{
"dataset_name": {
"pp-intel-dcache": "/path/to/intel_l1_dcache",
"pp-intel-icache": "/path/to/intel_l1_icache",
"pp-amd-dcache": "/path/to/amd_l1_dcache",
"pp-amd-icache": "/path/to/amd_l1_icache",
},
}
Then run
python pp_image.py --exp_name="CelebA_pp" --dataset="CelebA" --cpu="intel" --cache="dcache"
to approximate the manifold. To reconstruct unknonw images from the collected cache set accesses, uncomment
engine.load_model("/path/to/model.pth")
engine.inference(test_loader, "test")
in pp_image.py
. The reconstructed images will be saved in output/CelebA_pp/recons
. Follow the same procedure for other media data.
We release our trained models and all SC09 audios reconstructed from side channels collected by Prime+Probe on Intel L1 D cache.
7. Noise Resilience
We have the following noise insertion schemes (see more details in our paper).
Pin logged trace | Prime+Probe logged trace |
---|---|
Gaussian | Leave out |
Shifting | False hit/miss |
Removal | Wrong order |
To insert the "shifting" noise into pin logged trace, go to code
and run
python output_noise.py --exp_name="CelebA_cacheline" --dataset="CelebA" --side="cacheline" --noise_op="shift" --noise_k=100
Images reconstructed from noisy cache line records will be saved in output/CelebA_cacheline/recons_noise
by default.
To insert the "wrong order" noise into prime+probe logged trace, you need to modify code/pp_image.py
as
# test_dataset = RealSideDataset(args, split=args.data_path[args.dataset]["split"][1])
test_dataset = NoisyRealSideDataset(args, split=args.data_path[args.dataset]["split"][1])
and uncomment
engine.load_model("/path/to/model.pth")
engine.inference(test_loader, "test")
to reconstruct unknown images from noisy side channel records.
The procedure is same for other media data. Note that in order to assess the noise resilience, you should NOT approximate manifold (i.e., train the model) using the noisy side channel.
8. Customization
All parameters are set in code/params.py
. You can customize the hyper-parameters for approximating manifold.
All datasets are implemented with OOP manner in code/data_loader.py
. You can modify the dataset class to support your own data.
All models are also implemented with OOP manner in code/model.py
. You can build a new framework from new models.
Citation
TBA.
If you have any questions, feel free to contact with me ([email protected]).