AI-Based Detection and Classification of Adversarial and Fault-Induced Threats in Split Computing

Split Computing (SC) has emerged as an effective strategy for deploying Deep Neural Networks (DNNs) on edge and Internet of Things (IoT) platforms. By partitioning inference between constrained edge devices and powerful cloud servers. However, the intermediate feature representations exchanged between the Head and Tail models can be corrupted either by deliberate adversarial perturbations, obtained by crafting the input to the Head, or by unintended hardware faults affecting the computations in the Head. Both phenomena manifest as distortions in the feature maps, but they have distinct causes and require different responses. Therefore, detecting and classifying such anomalies is crucial for dependable and secure SC-based inference. This work formulates feature-map monitoring in SC as a three-class classification problem. A lightweight classifier is trained to distinguish among clean, fault-induced, and attackinduced feature maps at the split boundary. It thereby determines whether corruption is present and identifies its most likely source, enabling the detection and classification of threats. We conduct a detailed analysis of how hardware faults and adversarially crafted inputs in the Head affect the distribution of entries in the intermediate feature map. Performing a design-space exploration across multiple SC configurations. Experiments on SC ResNet-50 show that the SC configuration, which involves intermediate 3-channel feature maps, achieves an F1-score of 97.84% on the clean/adversarial/faulty classification task, with a 31.07% computational overhead relative to baseline SC inference. These results indicate that three-class feature-map classification is a practical and effective front-end for cause-aware analysis in SC deployments.