Dynamic Neural Network is All You Need: Understanding the Robustness of Dynamic Mechanisms in Neural Networks

Mirazul Haque, Wei Yang

Abstract: Deep Neural Networks (DNN) based solutions are being used to solve different day-to-day problems. Recently, DNNs are being deployed in real-time systems, and lowering the energy consumption and response time has become the need of the hour. To address this scenario, researchers have proposed early-exit Dynamic Neural Networks (DyNNs), where the computation is dynamic based on the input complexity. DyNNs are generally designed and based on larger static DNNs (SDNN). As the DyNNs decrease the energy consumption, it also becomes important to evaluate the robustness of DyNNs to ensure safety. However, there has not been a significant number of works focusing on the robustness of DyNNs. To address this issue, we propose systematic studies to evaluate the robustness of DyNNs. For that purpose, we propose four research questions. These studies are performed on three models and two datasets. Through the studies, we find that DyNNs are more robust than SDNNs, and DyNNs can be used to generate adversarial samples efficiently. We also provide insight into the design choices through research studies. Finally, we propose a novel attack that can decrease the effectiveness of the DyNNs and can be used to evaluate design choices in DyNN.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/arxiv:2308.08709/code)