Systematic Generation of Adversarial Datasets with Controllable Noise Levels
Management Strategies and Engineering Sciences · 2026 · 8 (5) · 1-9
DOI: 10.61838/msesj.345 ↗Kurz gefasst
Fachartikel in der Zeitschrift Management Strategies and Engineering Sciences, 2026. Verfasst von Mohammad Reza Norouzi.
Abstract (Originalsprache)
The rapid proliferation of user-generated textual content on social networks and digital platforms has created significant challenges for sentiment analysis systems. These challenges are more pronounced in the Persian language due to the scarcity of high-quality datasets, orthographic variability, and the high sensitivity of models to noise. One of the most critical issues is the vulnerability of machine learning models to textual noise and adversarial attacks, which can lead to substantial performance degradation. The objective of this study is to propose a systematic approach for generating adversarial textual datasets with controllable noise levels in order to evaluate and enhance the robustness of Persian sentiment analysis models. In this research, a baseline Persian sentiment analysis dataset was first preprocessed. Subsequently, a framework was designed to introduce targeted noise types, including word substitution, deletion, insertion, and permutation. For each type of noise, an intensity parameter was defined to enable precise control over the degree of perturbation. The adversarial data were generated independently of any specific model, and each instance was annotated not only with its sentiment label but also with metadata specifying the type and level of noise applied. The performance of several sentiment analysis models was then evaluated before and after training with the adversarial dataset. The results indicated that models trained exclusively on clean data experienced significant performance degradation when exposed to adversarial samples, particularly under substitution and deletion noise. In contrast, training with the generated adversarial dataset led to a considerable improvement in noise robustness and performance stability. The findings suggest that the systematic generation of adversarial data with controllable noise constitutes an effective instrument for sensitivity analysis and robustness enhancement in Persian sentiment analysis models and can play a critical role in the development of reliable systems under real-world conditions.
Zitieren
APA
Norouzi, M. R. (2026). Systematic Generation of Adversarial Datasets with Controllable Noise Levels. Management Strategies and Engineering Sciences, 8(5), 1-9. https://doi.org/10.61838/msesj.345
Vancouver
Norouzi MR. Systematic Generation of Adversarial Datasets with Controllable Noise Levels. Management Strategies and Engineering Sciences. 2026;8(5):1-9. doi:10.61838/msesj.345
BibTeX
@article{Norouzi2026,
author = {Norouzi, Mohammad Reza},
title = {Systematic Generation of Adversarial Datasets with Controllable Noise Levels},
journal = {Management Strategies and Engineering Sciences},
year = {2026},
volume = {8},
number = {5},
pages = {1--9},
doi = {10.61838/msesj.345}
}
EUVIMED