New Research by TSYS School Computer Scientist Riduan Abid Examines Use of Split Neural Networks in Edge Computing

Cloud computing is a critical component in the success of 5G and 6G networks, particularly given the computation-intensive nature of emerging applications. Despite all it advantages, cloud computing faces limitations in meeting the strict latency and bandwidth requirements of applications such as eHealth and automotive systems. To overcome these limitations, edge computing has emerged as a novel paradigm that bring computation closer to the user. Moreover, intelligent tasks related to deep learning demand more memory and processing power than edge devices can handle. To address these challenges, methods like quantization, pruning, and distributed inference have been proposed. A new study by TSYS School computer scientist Riduan Abid, Salmane Douch, Khalid Zine-Dine and Driss Bouzidi of Mohammed V University in Rabat, and Driss Benhaddou of Alfaisal University examines a promising approach for running deep learning models at the edge that employs split neural networks. Split neural networks feature a neural network architecture with multiple early exit points, allowing the model to make confident decisions at earlier layers without processing the entire network. This not only reduces memory and computational demands but it also makes split neural networks well-suited for edge computing applications. As the use of split neural networks expands, ensuring their safety—particularly their robustness to perturbations—becomes crucial for deployment in safety-critical scenarios. Abid et al.'s new study, set to appear in a future issue of IEEE Access, presents the first in-depth study on the robustness of split Edge Cloud neural networks. In doing so it reviews state-of-the-art robustness certification techniques and evaluate split neural networks robustness using the auto_LiRPA and Auto Attack libraries, comparing them to standard neural networks. The results demonstrate that split neural networks reduce average inference time by 75‘% and certify four to 10 times more images as robust, while improving overall robustness accuracy by 1% to 10%.