In-Progress: Structured Pruning in the Wild: Benchmarking Practical Robustness Under Real-World Corruptions

Convolutional Neural Networks (CNNs) are widely adopted in Internet-of-Things(IoT) applications, where real-time deployment on resource-constrained edge devices is critical. However, when these devices operate in uncontrolled environments, the input they process is often imperfect and subject to various types of corruption, such as noise, blur, or adversarial perturbations. While numerous structured pruning methods have demonstrated their ability to maintain model performance under benign conditions - while simultaneously providing computational efficiency and storage benefits - it remains unclear how well they perform when faced with real-world corrupted inputs. Do these pruning techniques maintain robustness in such challenging scenarios, or do they significantly degrade in performance? Furthermore, pruning methods designed to be adversarially robust claim to withstand certain targeted perturbations, but do they actually generalize? This pilot study aims to investigate these questions by systematically benchmarking several well-established structured pruning techniques applied to CNNs.