DeepAdversaries: Examining the Robustness of Deep Learning Models for Galaxy Morphology Classification

We present the data used in "DeepAdversaries: Examining the Robustness of Deep Learning Models for Galaxy Morphology Classification".

Data processing and analysis pipelines in cosmological survey experiments introduce data perturbations that can significantly degrade the performance of deep learning-based models. We consider perturbations associated with two primary sources:

1) increased observational noise as represented by higher levels of Poisson noise;

2) data processing noise incurred by steps such as image compression or telescope errors as represented by one-pixel adversarial attacks. 

We also test the efficacy of domain adaptation techniques in mitigating the perturbation-driven errors. Successful  development  and  implementation  of methods that increase model robustness in astronomical survey pipelines will help pave the way for many more uses of deep learning for astronomy.

To create our data, which emulates Vera C. Rubin LSST observations, we use GalSim (Rowe et al. 2015) and follow the same procedure as in Sanchez et al. (2021). We create two sets of survey-emulating images, by applying an exposure time directly to the raw images:

1) low-noise ten-year survey ("Y10'') - baseline data our models are trained on;

2) high-noise one-year survey ("Y1'') - noisy data.

We use one-pixel attack (Su et al. 2017) to attack a 150-image test set sub-sample of Y10 images, in order to emulate pixel-level perturbations, that might naturally arise in astronomical data pipelines.

Data:

     Images Y10: images_Y10_train.npy;   images_Y10_valid.npy;   images_Y10_test.npy

     Noisy mages Y1: images_Y1_train.npy;   images_Y1_valid.npy;   images_Y1_test.npy

     Labels valid for both Y10 and Y1: labels_train,npy;   labels_valid.npy;   labels_test.npy 

Small 150-image sub-samples:

    Images: images_Y10_test_150.npy;   images_Y1_test_150.npy

    Labels: labels_test_150.npy