This repository presents an academic study of data poisoning attacks in machine learning systems.
The research investigates training-time adversarial manipulations, their formal threat models, practical implementation, and impact on the security and robustness of machine learning models.
Data Poisoning Attacks, Adversarial Machine Learning, Machine Learning Security, Training-Time Attacks, Model Robustness.
The objectives of this research are:
- to formally define data poisoning attacks and their threat models;
- to analyze existing attack strategies and taxonomies;
- to evaluate the effect of poisoned training data on model performance;
- to reproduce selected poisoning attacks through controlled experiments;
- to review and compare existing defense and mitigation techniques.
This research focuses on:
- supervised learning under adversarial training conditions;
- label-flipping and clean-label poisoning attacks;
- targeted and untargeted poisoning scenarios;
- backdoor injection during the training phase;
- security evaluation and robustness analysis of machine learning models.
- Mikhail Shumakov
- Arseniy Shchekin