Adversarial susceptibility analysis for water quality prediction models

Area of Science:

• Environmental Science
• Public Health
• Artificial Intelligence

Background:

• Water quality is crucial for health and sustainability.
• Urbanization and industrialization increase waterborne disease risks.
• Pathogen detection in water sources is vital for disease prevention.

Purpose of the Study:

• Investigate water contamination patterns in Gujarat using machine learning.
• Identify key pathogens affecting water quality.
• Evaluate the robustness of AI models against data corruption.

Main Methods:

• Applied machine learning classifiers: HistGradientBoosting, Random Forest, AdaBoost, Bagging, Decision Tree, and LSTM.
• Utilized SHapley Additive exPlanations (SHAP) for feature interpretation.
• Assessed model resilience using adversarial perturbations (FGSM, PGD) and adversarial training.

Main Results:

• Random Forest and Bagging classifiers achieved 98.53% accuracy in predicting water quality.
• Adversarial attacks caused up to a 56% performance drop, reduced to 10% after adversarial training.
• Neural network models demonstrated greater resilience to adversarial attacks compared to traditional machine learning models.

Conclusions:

• Machine learning, particularly Random Forest and Bagging, shows high accuracy for water quality assessment.
• AI models require adversarial training to ensure reliability in real-world conditions with data noise.
• Developing resilient AI systems is essential for effective public health and water safety management.