Adversarial Reprogramming has demonstrated success in utilizing pre-trained
neural network classifiers for alternative classification tasks without
modification to the original network. An adversary in such an attack scenario
trains an additive contribution to the inputs to repurpose the neural network
for the new classification task. While this reprogramming approach works for
neural networks with a continuous input space such as that of images, it is not
directly applicable to neural networks trained for tasks such as text
classification, where the input space is discrete. Repurposing such
classification networks would require the attacker to learn an adversarial
program that maps inputs from one discrete space to the other. In this work, we
introduce a context-based vocabulary remapping model to reprogram neural
networks trained on a specific sequence classification task, for a new sequence
classification task desired by the adversary. We propose training procedures
for this adversarial program in both white-box and black-box settings. We
demonstrate the application of our model by adversarially repurposing various
text-classification models including LSTM, bi-directional LSTM and CNN for
alternate classification tasks.
