Simple explanation
SecAI+ doesn't require you to build models — it requires enough understanding of how they LEARN to reason about where that learning process can be attacked or misused. This lesson covers the foundational types.
Technical explanation
- Supervised learning — trains a model on labeled examples (input paired with a known correct output), so it can predict labels for new, unseen inputs. Most classification and prediction systems (spam filters, fraud detection) use this.
- Unsupervised learning — finds patterns or groupings in unlabeled data without being told the "correct" answer in advance. Clustering (grouping similar data points) is the classic example.
- Reinforcement learning — trains a model through trial and error, receiving reward or penalty feedback based on actions taken, rather than from labeled examples. Used in systems that need to make sequential decisions, like game-playing AI.
- Deep learning — uses neural networks with many layers to learn increasingly abstract representations of data; the foundation of most modern generative AI, including LLMs.
- Training from scratch vs. fine-tuning — training from scratch builds a model's parameters from random initialization (extremely resource-intensive); fine-tuning takes an already-trained model and further trains it on a smaller, narrower dataset to specialize it for a specific task, at a fraction of the cost.
Synonyms / related terms
| Term | Means | |---|---| | ML | Machine Learning | | Fine-tuning | Specializing a pre-trained model on new, narrower data | | Clustering | The classic unsupervised learning use case |
Concept Check
"A company wants to build a fraud-detection model using millions of past transactions, each already labeled as 'fraudulent' or 'legitimate.'" This is supervised learning, not unsupervised — the defining feature is that the training data already has known, correct labels attached, which the model learns to predict for new transactions.
Interview-style Q&A
Q: Why does a security practitioner need to understand these ML types, rather than leaving it entirely to data scientists? A: "Because the attack surface differs by type. A supervised model's attack surface centers heavily on its labeled training data — poison the labels, poison the model. An unsupervised model can be manipulated by feeding it data designed to shift what it considers a normal cluster. Understanding the type shapes what threat you actually need to defend against."
Memory trick
"Supervised has a Teacher, Unsupervised finds Patterns, Reinforcement earns Rewards" — three ML types, each tied to its defining learning mechanism.