Impact of Data Poisoning on Machine Learning Models: A Case Study on S&P 500 Index
10th International Conference on Computer Science and Engineering, UBMK 2025, İstanbul, Turkey, 17 - 21 September 2025, pp.172-178, (Full Text)
- Publication Type: Conference Paper / Full Text
- Doi Number: 10.1109/ubmk67458.2025.11206907
- City: İstanbul
- Country: Turkey
- Page Numbers: pp.172-178
- Keywords: AI security, data poisoning, data security, machine learning, S&P500 index movement, stock market movement
- TED University Affiliated: Yes
Abstract
The integration of ML models into critical domains such as finance, healthcare, and security-sensitive applications has raised safety concerns and increased the need to ensure their robustness against adversarial threats. One of the most concerning threats is data poisoning, where attackers manipulate the training data to degrade model performance or force specific outcomes. This paper examines how label flipping impacts the performance of machine learning models, specifically the Decision Tree, Random Forest, XGBoost, and SVM classifiers. Throughout experiments, we simulated label flipping in the training dataset and evaluated how these changes affected the model's outcomes. Models' resilience to poisoned data was measured using key matrices, including precision, recall, accuracy, and F1 score. The results indicate that when using poisoned data, Random Forest and XGBoost perform more robustly than Decision Tree and Support Vector Machine. This highlights the relative robustness of ensemble models (Random Forest and XGBoost) compared to single estimators (Decision Tree and SVM). This is attributed to the ensemble learning mechanisms that enable better generalization and error correction. This study provides a perspective on model robustness under adversarial conditions and emphasises the importance of algorithm selection while deploying ML systems where data integrity cannot be guaranteed.