Data contamination detection and defense algorithms for IoT edge computing

• Miao Liu,
• Maojia Lv,
• Yaru Guo,

Vol. 19, No. 6, June 30, 2025
10.3837/tiis.2025.06.013, Download Paper (Free):

• Edge Computing
• Data Enhancement
• data pollution
• Deep Learning
• triadic loss

Abstract

Edge Computing (EC) as the essence is that the cloud server distributes computing tasks to the node servers linked to it. The complex environment awareness supported by server cluster computing driven by big data is a hot research and infrastructure focus for the current and future period. To address the challenge of limited computational resources in edge nodes within edge computing systems, as well as the difficulty in accurately detecting service failures triggered by malicious attacks from other edge nodes, this paper proposes an improved Efficient edge computing Secure Deep Learning (DL) algorithm (Efficient edge computing Secure Deep Learning Method Based on Data Pollution, DL) for data pollution. This paper proposes an improved Efficient edge computing Secure Deep Learning Method Based on Data Pollution (EMD) algorithm to solve the problem of fluctuating and inaccurate computing results of edge nodes due to their poor robustness, data distortion or mild qualitative changes. The network is trained by adding Gaussian Noise (GN) to the randomly selected batch samples, using the expanded data set, and by incorporating the newly introduced compound loss function from this paper, the challenging-to-optimize triplet method can be effectively trained. This enables the network to have a broader data fitting and prediction capability, and this addresses the issue of systemic failure, which arises when executing accurate operations at the edge nodes becomes challenging due to significant data corruption. In addition, an accelerated convergence function is proposed in this paper to accelerate the convergence speed of network training. The algorithm's precision improved by 2.56%, 4.26%, and 6.83% on the Cifar10 dataset compared to the baseline algorithm. On the 20% contaminated Cifar10 dataset, its precision increased by 2.43%, 4.24%, and 6.93% relative to the comparison algorithm. The results of the simulation indicate that the algorithm proposed in this paper is more effective at detecting both noisy-contaminated samples and those with random labeling errors, and the algorithm can achieve the best results within the specified training batches.

Statistics

Cite this article