Hidden Activations Are Not Enough:
This repository contains a very simple version of the code used to run the experiments in the paper titled Hidden Activations Are Not Enough: A General Approach to Neural Network Predictions, by Samuel Leblanc, Aiky Rasolomanana, and Marco Armenta. The paper can be found on arXiv. The repository for the code used in the paper can be found here.
Important: This code was not used to run the experiments of the aforementioned paper. To reproduce the experiments, please refer to the repository mentioned above.
The code being very simple, the file example.py contains everything one needs to understand to be able to run an experiment.
The algorithm mentioned in the Discussion section of the paper for the detection of out-of-distribution data is implemented.
To run it, you could for instance add the line exp.reject_predicted_out_of_dist(std_z=1, std_e=1.5, eps=0.1, eps_p=0.1, delta=2, delta_p=2) at the end of the file example.py.
The variable t^ε of the paper refers to std_z and ε = eps, ε' = eps_p.
Refering to the Discussion section, t^δ = std_e, δ = delta, and δ' = delta_p.
By changing dataset="mnist" to dataset="cifar10" in example.py, you can train an MLP on CIFAR-10, produce the matrices, and detect adversarial examples!
MIT