Definition
Data Poisoning in Machine Learning (ML) refers to attacks carried out manipulating training data to alter the learning process and eventually impacting on ML models’ inference.
Background
Learning from data allows ML models to solve a wide variety of problems with unprecedented performance. However, learning from data also implies that any issue affecting data may potentially influence the learning process and eventually the ML model performance. Attackers may try to exploit the learning process vulnerabilities to alter the ML model training outcome. More specifically, in a Data Poisoning attack, training data manipulation is considered. Data Poisoning highlights the tight relationship between ML security, data governance, and protection: a connection that is not limited to the technical side but that also reverberates, as an example, in the European Union Artificial Intelligence Act (or AI...
