SecureCovid

Implementation of secure covid project described in proposal.

In this project, our goal is to assess the effectiveness of differential privacy, a framework for protecting privacy of ML model.

can find most of the examples in notebook

Run Demo

Simply cd demo ; sh run_demo.sh, it will display a cxr image and the model output if you have the path defined correctly.

Shadow Model

To train, first prepare for the image training data and run

python shadow_train.py --data_path path/to/file

usage: shadow_train.py [-h] [--data_path DATA_PATH] [--out_path OUT_PATH]
                       [--weight_path WEIGHT_PATH] [--mode MODE]
                       [--model MODEL] [--valid_size VALID_SIZE]
                       [--learning_rate LEARNING_RATE] [--step_size STEP_SIZE]
                       [--gamma GAMMA] [--epoch EPOCH] [--name NAME]

Secure Covid Shadow Train

optional arguments:
  -h, --help            show this help message and exit
  --data_path DATA_PATH
                        Path to store the data
  --out_path OUT_PATH   Path to store the trained model
  --weight_path WEIGHT_PATH
                        Path to load the trained model
  --mode MODE           Select whether to train, evaluate, inference the model
  --model MODEL         Select which model to use
  --valid_size VALID_SIZE
                        Proportion of data used as validation set
  --learning_rate LEARNING_RATE
                        Default learning rate
  --step_size STEP_SIZE
                        Default step size
  --gamma GAMMA         Default gamma
  --epoch EPOCH         epoch number
  --name NAME           Name of the model

Attack Model

To train, first prepare for the input data and output data and run

python attack_train.py

sage: attack_train.py [-h] [--input_path INPUT_PATH]
                       [--target_path TARGET_PATH] [--out_path OUT_PATH]
                       [--weight_path WEIGHT_PATH] [--mode MODE]
                       [--label LABEL] [--valid_size VALID_SIZE]
                       [--learning_rate LEARNING_RATE] [--epoch EPOCH]
                       [--name NAME]

Secure Covid

optional arguments:
  -h, --help            show this help message and exit
  --input_path INPUT_PATH
                        Path to store the data
  --target_path TARGET_PATH
                        Path to store the data
  --out_path OUT_PATH   Path to store the trained model
  --weight_path WEIGHT_PATH
                        Path to load the trained model
  --mode MODE           Select whether to train, evaluate, inference the model
  --label LABEL         Select the label for the attack model
  --valid_size VALID_SIZE
                        Proportion of data used as validation set
  --learning_rate LEARNING_RATE
                        Default learning rate
  --epoch EPOCH         epoch number
  --name NAME           Name of the model

Test

Test Set

Out-of-Training	In-Training	Total
400	400	800

Shadow = 1, No Differential Privacy

	Accuracy	Loss
Positive	0.5241	0.8337
Negative	0.5509	0.7522

Shadow = 5, No Differential Privacy

	Accuracy	Loss
Positive	0.5293	0.7356
Negative	0.5106	0.7891

---

Shadow = 1, Differential Privacy

	Accuracy	Loss
Positive	0.4998	0.8170
Negative	0.5092	0.7825

Shadow = 5, Differential Privacy

	Accuracy	Loss
Positive	0.5020	0.7553
Negative	0.4925	0.7980