Membership Inference Attack with LGBM Model

Membership inference attack on neural networks with MNIST and CIFAR10 datasets
parser = argparse.ArgumentParser()

# inference type
parser.add_argument('--is_target_train', type=bool, default=True, help="Whether training target model")    
parser.add_argument('--is_DP', type=bool, default=True, help="Whether training target model with DP")    

# path:
parser.add_argument('--model_target_dir', type=str, default='.')
parser.add_argument('--data_train_dir', type=str, default='.')
parser.add_argument('--data_val_dir', type=str, default='.')
parser.add_argument('--data_train_shadow_dir', type=str, default='.')
parser.add_argument('--data_val_shadow_dir', type=str, default='.')            

# general:
parser.add_argument('--seed', type=int, default=0)
parser.add_argument('--train_target_size', type=int, default=2500, help='2 500, 5 000, 10 000, 15 000 // 4 600, 10 520, 19 920, 29 540')
parser.add_argument('--test_target_size', type=int, default=1000)
parser.add_argument('--number_shadow_model', type=int, default=25, help='25 50 MNIST and 100 cifar')

# learning:
parser.add_argument('--batch_size', type=int, default=10)
parser.add_argument('--learning_rate', type=float, default=0.001)
parser.add_argument('--momentum', type=float, default=0.5)
parser.add_argument('--epochs', type=int, default=25)
parser.add_argument('--decrease_lr_factor', type=float, default=1e-7)
parser.add_argument('--decrease_lr_every', type=int, default=1)

parser.add_argument('--reg_lambd', type=int, default=10)
parser.add_argument('--n_estimators', type=int, default=10000)
    
# statistics:
parser.add_argument('--dataset', type=str, default='MNIST', help="MNIST, CIFAR10, CIFAR100")
parser.add_argument('--training_size_value', type=list, default=[4600, 10520, 19920, 29540], help="[2500, 5000, 10000, 15000]")
parser.add_argument('--number_shadow_value', type=list, default=[100], help='[2, 10 ,20, 50, 100]')
parser.add_argument('--epoch_value', type=list, default=[100], help="[2, 10 ,20, 50, 100]")
parser.add_argument('--experments', type=int, default=3)
    
# differential privacy
parser.add_argument('--eps', type=int, default=2)    
parser.add_argument('--set_eps', type=list, default=[1,20,100])    

args = parser.parse_args()


# Example 1: training target model with various DP levels
if args.is_target_train:
    
    acc_eps = []
    
    for args.eps in args.set_eps:
        acc_list = []
        for i in range(args.experments):
        
            precision_general, recall_general, accuracy_general, precision_per_class,\
                recall_per_class, accuracy_per_class = mi(args)
            acc_list.append(accuracy_general)
            print('\nEps:', args.eps, 'Attack acc:', accuracy_general)    
        acc_eps.append(sum(acc_list)/len(acc_list))
    print('\nAttack acc list:', acc_eps)
    
# Example 2: attack the trained model with allocated datasets   
else:

    set_idx = [0,0,0,0]
    set_eps = [2,5,10,20]
    
    acc_eps = []

    for i in range(len(set_idx)):
        args.model_target_dir = './mi_attack/model-client{}_eps{}_delta0.001.ckpt'.format(set_idx[i], set_eps[i])
        args.data_train_dir = './mi_attack/dict_train-mnist.pt'
        args.data_val_dir = './mi_attack/dict_test-mnist.pt'
        args.data_train_shadow_dir = './mi_attack/dict_train_shadow-mnist.pt'
        args.data_val_shadow_dir = './mi_attack/dict_test_shadow-mnist.pt'
    
        acc_list = []
        for k in range(args.experments):
            precision_general, recall_general, accuracy_general, precision_per_class,\
                recall_per_class, accuracy_per_class = mi(args)
            acc_list.append(accuracy_general)
            print('\nEps:', set_eps[i],'Accuracy:', accuracy_general)
        acc_eps.append(sum(acc_list)/len(acc_list))
    print('\nAttack acc list:', acc_eps)
AdamWei-boop/Membership-Inference-Attack-LGBM

Membership Inference Attack with LGBM Model
