SecureV2X

SecureV2X is a research framework for securely computing V2X applications. Namely, we design secure protocols for private object detection and red light violation detection with YOLOv5n, and drowsiness detection via CompactCNN. To use our framework, please follow these steps:

1. First, install crypten by running

   pip install crypten
   

   CrypTen is not currently available through Conda and so pip must be used to install it. Our framework relies upon the use of secure primitives implemented by CrypTen and so this package will be necessary to run FastSec-YOLO or CryptoDrowsy. Additionally, please make sure to install all of the packages listed in requirements.txt.

2. Once crypten is installed, run the following in python

   import crypten
   crypten.__file__

   This will output the location of ../crypten/__init__.py. Navigate to the nn folder from there, and replace the default version of crypten/nn/module.py with the module.py file provided at case_studies/module.py from our repository. Our version provides the necessary protocols to run SecureV2X. NOTE: SecureV2X will not work correctly if this step is not followed correctly. Please take care to correctly locate the original module.py file and replace it with the updated code

Next, run pip install requirements.txt to install the required packages for SecureV2X. The most recent version of each should work with our code.

Scalability Simulations

We evaluate the scalability of SecureV2X to a network of clients by running multiple instances of CryptoDrowsy and FastSec-YOLO simultaneously. These simulations are conducted in agent_sim.py and will require the following steps to execute properly.

1. Navigate to this repository and download y2.mp4 directly from the videos folder. Store this video file in case_studies to perform multi-client secure inference.
2. Run python3 agent_sim.py y2.mp4 in case_studies to run multi-client inference.
3. check the output results in experiments/agg_results.csv where
   • exp = (x-y) pairs where x is the number of cryptodrowsy clients run and y is the number of fastsec-yolo clients run
   • d_inf_time = time required for each CryptoDrowsy inference (on average)
   • r_inf_time = time required for each FastSec-YOLO client inference (on average) from a batch of 4 video frames from y2.mp4

Note the output results correspond to section 5.3 from the associated paper.

CryptoDrowsy

Three core results are provided for CryptoDrowsy - namely running time (over the CPU) in addition to communication cost and rounds required per inference. To reproduce the results we provide in the paper, please navigate to case_studies and run python3 run_cryptodrowsy. This will perform inference over each of the 314 (3-second 128Hz) eeg samples corresponding to subject 9 as studied in this paper. The accuracy, communication, round complexity, and timing results will be reported to the CLI. As was done in the prior work, we have pretrained a model for each of the 11 test subjects, and have included their respective labels and features as well. run_cryptodrowsy.py sets subject 9 as the default. However, if you would like to run inference for any one of the other 10 subjects, simply change sub9 to subx where $x$ is any number from 1 to 11 (inclusive). Likewise, change 9-drowsy_features.pth and 9-drowsy_labels.pth to x-drowsy_features.pth and x-drowsy_labels.pth where x is the same as for subx.

In order to reproduce private inference over CompactCNN using Delphi, first, install rust and then run the following commands:

sudo apt install cmake pkg-config g++ gcc libssl-dev libclang-dev
rustup install nightly

Next, navigate to delphi/rust/experiments/src/validation/compactCNN/validation and run python3 run_validation.py. If an error occurs, please carefully check to make sure that all of the paths listed in run_validation.py are correct. Before running this functionality

• weights_path - case_studies/driverdrowsiness/pretrained/sub9/model.npy
• eeg_test_data_path - folder containing all of the .npy data samples for CompactCNN
  • default value - delphi/rust/experiments/src/validation/compactCNN/Eeg_Samples_and_Validation given as a relative path from delphi/rust/experiments/src/inference
• num_samples - max of 314 - recommend a lower number - all 314 may take at least 8 hours to run all 314 samples.
• accuracy_results_path - relative path from delphi/rust/experiments/src/inference to delphi/rust/experiments/src/validation/compactCNN/Eeg_Samples_and_Validation/Classification_ResultsX.txt
• output_file - will divert text output of delphi run to a saved text file. By default, this is set as the relative path to delphi/rust/experiments/src/validation/compactCNN/validation_runs/vlaidation_runX.txt from delphi/rust/experiments/src/inference/

Don't be alarmed if a connection refused error occurs during the run. Simply restart the process. If you intend to run all 314 samples, you will want to check delphi/rust/experiments/src/inference/compactCNN/sequential_inference.rs to adjust values for the range of samples to run if you have already run several by the time a connection refused error occurs.

In order to run CompactCNN on CrypTFlow2, you will need to clone EzPC and use the custom compiler they provide to compile a secure version of CompactCNN. Follow the instructions provided for running SCI at the link.

FastSec-YOLO

FastSec-YOLO can be run by doing the following: First, navigate to case_studies and run run_fastsec_yolo.py with a command line argument of the form source/image.jpg where image.jpg is the image you want to perform secure object detection over. If you would like to perform inference over your own image, please upload that image to the source folder directly. Otherwise, you may test this functionality by running inference over the bus.jpg image which is held within the source folder.

Once inference is performed, you may check the bounding box detection results in the experiments/img_visuals folder which will have been generated. The most likely issue to occur during this step is the improper initialization of CUDA. If this happens, please check to make sure that your cuda settings are properly set. ⚠️ The code should try to handle potential issues automatically, but please be aware

Experiments

First, download the coco128 data from this link. If the link does not work for some reason, please visit ultralytics/yolov5 on github and download coco128.zip from release v1.0. Make a note of the absolute path to the image and label folders on your system, or move the image and label data into the repository so that you can Our experiments for FastSec-YOLO can be reproduced as follows:

1. Navigate to case_studies/traffic_rule_violation
2. Run python3 fastsec_yolo_val.py benchmark <labels_folder> <images_folder> where <labels_folder> contains the labels for the coco128 train image set, and <images_folder> contains the actual images for the coco128 image dataset.

Image and batch size experiments can be reproduced by chnaging "benchmark" to either "batch_size" or "img_size" for batch and image size experiments respectively. The results of this experiment will be output to case_studies/traffic_rule_violation/experiments/model_type_exps where each file in that folder is a .csv file containing the results of both plaintext and secure inference over coco128 for each model (yolov5: n, s, m, l, x). ❗Note that results may differ from those reported in the associated paper due to computational machinery differences.