Input-Slice Attacks and Defense against Pseudo-Model-Parallelism-Based Collaborative DNN

Authors:

Abstract
Collaborative deep neural network (CDNN) inference across distributed edge devices provides scalability and resilience in mission-critical scenarios. Prior research has concentrated on adversarial attacks targeting intermediate feature maps within pseudo-model-parallelism Convolutional Neural Networks (ps-CDNNs), where feature-level disturbances are relatively localized and statistically detectable. In this paper, we introduce and examine a new type of input-level slice attack, in which adversaries manipulate input image segments before entering the network pipeline. We show that such attacks are much harder to detect with traditional statistical or filter-based defenses because of their subtler effects on aggregated feature distributions. To combat this, we propose an ensemble of lightweight machine learning detectors that utilize statistical signatures from feature maps, achieving an F1 score of up to 87% while keeping a low inference overhead. Our comparative analysis across attack surfaces (input vs. intermediate) and model architectures (VGG16, ResNet18) indicates that input-level perturbations not only bypass classic defenses but also more effectively degrade accuracy in multislice settings. These findings reveal a previously underexplored vulnerability in collaborative inference frameworks and emphasize the need for adaptive, model-agnostic detection strategies at the system edge to secure artificial intelligence (AI) models.