th-b/linear_filtration

Linear filtration

The file filtration.py contains an implementation of linear filtration as described in Machine Unlearning: Linear Filtration for Logit-based Classifiers. Please cite this paper if you are going to use the code in this repository.

The function filtration_matrix() calculates the filtration matrix. Its arguments are:

• predictions sample used to estimate mean predictions.
• labels i-th entry is the true class label of the i-th prediction.
• num_classes number of classes (before filtration).
• filter_classes array of class labels to unlearn, e.g. [3,5] to unlearn classes 3 and 5.
• mode choose from 'naive', 'normalization', 'randomization', or 'zeroing'.

The function make_filtered_model() applys filtration to a model. Its arguments are:

• model a Keras model to apply filtration to.
• data, labels, num_classes, filter_classes, mode see description of filtration_matrix().
• lin_index the index of the last dense layer in your network. Note that if you use tf.keras.layers.Activation('softmax') Keras counts it as its own layer, so lin_index should be -2 in this case. If you use tf.keras.layers.Dense(n, activation='softmax') lin_index should be -1.
• apply_softmax whether to apply softmax activations to the filtered model.

Example: model inversion for AT&T Faces

The file att_faces_example.ipynb contains an example where we apply linear filtration a model trained on the AT&T Laboratories Cambridge Database of Faces, to investigate its effect on reconstructions of face images by gradient ascent ("model inversion").