Combining randomized smoothing with existing classifier/model

[DmitriiGudin]

Hello friends,

I am working on a course project in which I am trying to combine the XGBoost classifier (trained, for instance, on the MNIST data) with randomized smoothing to achieve provable certifiable robustness. I am new to ART and am having trouble figuring out the syntax.

I have a base trained classifier clf obtained as follows:

def train_base_model (train_X, train_Y, test_X, test_Y): # Trains and returns an XGBoost model. train_X must be an array of 1-dimensional arrays, and train_Y is an array of integers corresponding to different labels.
    parameters = {"objective": "multi:softprob", "metric": "accuracy", "num_class": len(params.classes)}
    dtrain = xgb.DMatrix(train_X, np.argmax(train_Y, axis=1))
    dtest = xgb.DMatrix(test_X, np.argmax(test_Y, axis=1))
    evals = [(dtest, "test"), (dtrain, "train")]
    model = xgb.train(params=parameters, dtrain=dtrain, num_boost_round=50, evals=evals)
    clf = XGBoostClassifier(model=model, clip_values=(min_pixel_value, max_pixel_value), nb_features=params.N_features, nb_classes=len(params.classes))
    return clf

Now, I tried using the RandomizedSmoothingMixin wrapper from art.estimators.certification.randomized_smoothing to combine it either with the classifier clf, or the base model model, from the function above - but could not figure out how to do it. Reading this page still left me confused.

Any tips are appreciated. Thanks!

[beat-buesser]

Hi @DmitriiGudin I think it's an interesting project. You would have to create a similar estimator as for example PyTorchRandomizedSmoothing but for XGBoost. This file also shows how to use RandomizedSmoothingMixin.