AI Model Governance: Bias, Explainability and Security Review
Security teams are increasingly asked to sign off on machine learning systems, and the questions that matter are not the ones most people expect. Whether the model is accurate is the data scientist's problem. Whether the model is governable — inventoried, explainable, monitored and reversible — is yours. Bias sits at the center of that, and it is worth being precise about why. A biased model is not merely unfair; it is a model whose behavior you cannot predict from its specification. That is a security property. Where Bias Comes From Three sources appear repeatedly, and exams distinguish between them. Sampling bias means the training data does not represent the population the model will face. An intrusion detection model trained only on one company's network learns that company's normal, not normal in general. Label bias means the ground truth itself encodes past human judgment. If historical alerts were triaged by analysts who consistently escalated traffic fr...