Harmonization for a black-box deep learning model

Motivation: Most commercially available deep-learning tools are black-box models, where additional training is not possible. This may hamper the performance for unseen-data. This work proposes novel harmonization concept for this black-box model. Goal(s): We propose a harmonization pipeline, BboxHarmony, that trains a harmonization network for a black-box model. Approach: First, randomly perturbed images were evaluated for the black-box model and the images with high performance results (pseudo-target domain images) were used to train the harmonization network. Then, the network was refined for black-box model using zeroth-order optimization that approximates backpropagation. Results: BboxHarmony successfully created harmonized images that provided high performance for the black-box model. Impact: BboxHarmony proposes a novel concept of harmonizing data for a black-box model and may have an important impact in the real-world where most commercial networks are black-box.