Robust detection of feature-space adversarial attacks in distributed edge inference using cascade Mahalanobis models

Artificial Intelligence of Things (AIoT) applications rely on distributed edge inference, where models are partitioned across edge devices and edge cloud nodes to meet latency and resource constraints. This exposes intermediate feature representations transmitted over IoT networks, creating vulnerabilities at partition boundaries. Detecting such feature-space attacks is challenging because boundary tensors are also affected by quantization, communication noise, and device variability. This paper proposes a lightweight statistical monitoring framework for securing distributed inference based on Mahalanobis distance modeling of clean boundary feature statistics. Two complementary detectors are developed: a computationally efficient diagonal Mahalanobis detector for fast anomaly screening, and a robust variant with feature normalization and covariance shrinkage for improved reliability under distortions. These detectors are combined in a cascade design where the first stage filters easy samples and the second stage refines uncertain cases. The framework is evaluated under channel-sparse black-box perturbations and realistic deployment impairments including quantization, additive noise, and feature dropout. Experiments across multiple architectures demonstrate that the proposed approach maintains controlled false-positive rates while achieving strong detection performance against optimized attacks and under non-ideal transmission conditions. Additional evaluations using MobileNetV3-Large, EfficientNet-Lite0, and ViT-Tiny extend the analysis to lightweight convolutional and transformer-based architectures and examine the sensitivity of the proposed framework to architecture family and collaboration boundary.