Machine Learning Model Data Poisoning Attacks and Countermeasures Research - Defensive Method Using Dynamic Network Structure Adjustment and Adaptive Learning Weights

Aug 3, 2026

Machine learning models are widely adopted in NLP, image and video processing, yet data poisoning attacks undermine performance and reliability by maliciously tampering with training data. The research focuses on building an effective defense mechanism to enhance robustness. The proposed method uses dynamic network structure adjustment and adaptive learning weights, improving stability of the decision boundary while adapting to shifting data distributions. Experiments validate improved accuracy under attack, with reduced degradation and stronger resilience in complex environments.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages8

Reading Time20 min

AccessFree to download

Introduction

Related work and limitations

Purpose and contributions

Data and methods

Data sources

Data preprocessing and normalization

Machine Learning Model Data Poisoning Attacks and Countermeasures Research Ruoheng xi School of Computer Engineering and Science, Shanghai University, Shanghai, 200444, China Abstract. Currently, machine learning models are widely used in natural language processing, image and video processing, and other fields. However, security issues such as data poisoning attacks threaten their performance and reliability. It is of great significance to study the defense mechanism. Therefore, this paper proposes a defense method based on dynamic network structure adjustment and adaptive learning weights and analyzes its effect in improving the robustness of model poisoning attacks. Experiments show that the accuracy of the model under poisoning attack drops from 100% to 76.67%, and the accuracy increases to 83.33% after adopting the defense mechanism. By dynamically adjusting the network structure and learning weights, the model can adapt to changes in data distribution, reduce the interference of poisoning data, and stabilize the decision boundary without obvious underfitting. This study provides an effective reference for improving the robustness of the machine learning model under data poisoning attacks, which helps to ensure the stable operation of the model in complex environments. 1 Introduction With the wide application of machine learning technology in key fields such as healthcare, transportation, and finance, its security problem has gradually become a research hotspot. The training and deployment of machine learning models are highly dependent on data, however, data may face various security threats in the process of collection, storage and use, among which data poisoning attacks are particularly prominent. Data poisoning attacks interfere with the model training process by maliciously tampering with the training data, thus reducing the performance and reliability of the model, and may even trigger risks such as data leakage. Therefore, it is of great theoretical and practical significance to study how to effectively defend against data poisoning attacks and guarantee the robustness and stability of machine learning models[1] . Currently, the defense methods for data poisoning attacks mainly focus on data preprocessing, model robustness optimization and anomaly detection. For example, Ma et al. [2] studied data poisoning attacks and their defense methods in differential privacy learning, proposed a game model between attackers and defenders, and improved the robustness of the model by optimizing the defense strategy. [3] Müller et al. conducted a study on data poisoning attacks in regression learning, and proposed a defense mechanism [22122827@shu.edu.cn](22122827@shu.edu.cn) © The Authors, published by EDP Sciences. This is an open access article distributed under the terms of the Creative Commons Attribution License 4.0 ([https://creativecommons.org/licenses/by/4.0/](https://creativecommons.org/licenses/by/4.0/)). based on statistical detection, which is able to identify and filter out some of the malicious data[4] . In addition, Chen et al. proposed an attack-independent defense method (De-Pois) to reduce the impact of poisoned data by optimizing the training process of the model[5] . However, these methods still have some limitations in practical applications, for example, data cleaning techniques make it difficult to completely identify malicious samples hidden in normal data, while model robustness optimization methods may increase model complexity and computational cost. Therefore, designing an efficient and reliable defense mechanism under the premise of guaranteeing model performance has become a key issue to be solved. The purpose of this paper is to propose a defense mechanism based on dynamic network structure adjustment and adaptive learning weights to improve the robustness of machine learning models under data poisoning attacks. This paper first introduces the knowledge of data poisoning at

FAQs

What is the main threat addressed in this research?

The paper addresses data poisoning attacks that maliciously tamper with training data, degrading model performance, reliability, and potentially causing additional risks such as data leakage.

What defense method does the paper propose?

It proposes a defense mechanism based on dynamic network structure adjustment and adaptive learning weights to improve robustness against model poisoning attacks.

How is the effectiveness of the defense mechanism evaluated?

Effectiveness is demonstrated through experiments on the Iris dataset, showing reduced accuracy drop under poisoning attacks and improved accuracy after applying the defense.

Aug 3, 2026

Upload by Genevieve