Training Set Attack

Training set attacks involve injecting malicious data points into a machine learning model's training data to manipulate its behavior, often targeting specific test instances or biasing overall predictions. Current research focuses on developing both more effective attack strategies, such as those leveraging co-training or game-theoretic approaches, and robust defense mechanisms, including improved influence estimation and dimensionality reduction techniques. Understanding and mitigating these attacks is crucial for ensuring the reliability and security of machine learning systems across diverse applications, from recommender systems to general classification tasks. The ability to identify targeted attacks and their victims is a particularly active area of investigation.