Toward Adversary-aware Non-iterative Model Pruning through <u>D</u> ynamic <u>N</u> etwork <u>R</u> ewiring of DNNs

We present a dynamic network rewiring (DNR) method to generate pruned deep neural (DNN) models that both are robust against adversarially generated images and maintain high accuracy on clean images. In particular, the disclosed DNR training is based unified constrained optimization formulation using novel hybrid loss function merges sparse learning with adversarial training. This strategy dynamically adjusts inter-layer connectivity per-layer normalized momentum computed from function. To further improve robustness of models, we propose DNR++, an extension where introduce idea parametric Gaussian noise tensor added weight tensors yield regularization. contrast existing pruning frameworks require multiple iterations, proposed DNR++ achieve overall target ratio only single iteration can be tuned support irregular structured channel pruning. demonstrate efficacy under no-increased-training-time “free” scenario, finally FDNR++, simple yet effective modification compressed requiring time comparable unpruned non-adversarial evaluate merits our methods, experiments were performed two widely accepted namely VGG16 ResNet18, CIFAR-10 CIFAR-100 as well Tiny-ImageNet. Compared baseline uncompressed methods provide over 20× compression all datasets without any significant drop either or classification performance. Moreover, extensive show consistently find better image performance than what achievable through state-of-the-art alternatives. insightful observations help make various model, parameter density, prune-type selection choices have open-sourced saved test codes ensure reproducibility results.