Dec
10
Paper accepted at EDCC 2026
December 10, 2025 | | Comments Off on Paper accepted at EDCC 2026
The paper entitled “CNN Robustness through Quantisation in Edge Devices: An Incomplete Solution” authored by Juan-Carlos Baraza-Calvo, David de Andrรฉs and Juan Carlos Ruiz Garcรญa has been accepted at EDCC 2026.
Abstract:
Deep learning has become a cornerstone for image analysis and object recognition. However, deploying Convolutional Neural Networks (CNNs) on embedded and IoT devices remains challenging due to strict energy and memory constraints. To address this gap, quantisation and hardware acceleration have emerged as complementary techniques that enable efficient CNN inference in resource-constrained environments. Beyond the efficiency benefits of combining both approachesโreduced memory footprint, faster inference, and lower energy consumption per operation โ several studies have also shown that quantisation enhances robustness against faults affecting network parameters. This practical study evaluates the behaviour of three state-of the-art quantised CNNs (ShuffleNet V2 0.5x, GoogLeNet, and Inception V3) under single and multiple bit-flip faults. While the results confirm the overall robustness of quantised CNNs, they also reveal the high sensitivity of a reduced set of parameters to injected faults, whose alteration can systematically compromise the inference process, even in the presence of a single or small number of bit-flips. This issue cannot be overlooked, especially when CNNs are deployed in devices or embedded systems operating in harsh environments. The precise identification of such fault-sensitive parameters is essential, as it paves the way for the selective deployment of fault-tolerance mechanisms, thereby improving CNN robustness in edge devices at lower cost and design effort.
Comments