---
id: CP-502
slug: supervised-unsupervised-reinforcement-learning
lang: en
title: Supervised, unsupervised and reinforcement learning
description: "How an AI system learns: labeled examples, unlabeled data, trial and reward. And why training data matters."
canonical_url: https://startbrain.io/cyberpedia/en/articles/supervised-unsupervised-reinforcement-learning/
markdown_url: https://startbrain.io/cyberpedia/en/articles/supervised-unsupervised-reinforcement-learning.md
publisher: Brain (https://startbrain.io)
license: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
cite_as: Brain, "Supervised, unsupervised and reinforcement learning", Cyberpedia, 2026. https://startbrain.io/cyberpedia/en/articles/supervised-unsupervised-reinforcement-learning/ (CC BY 4.0)
wing: 4
theme: understand_ai
reflex: air-002
schema: G3
updated: "2026-11-14"
translations:
  en: https://startbrain.io/cyberpedia/en/articles/supervised-unsupervised-reinforcement-learning/
  fr: https://startbrain.io/cyberpedia/fr/articles/apprentissage-supervise-non-supervise-renforcement/
  de: https://startbrain.io/cyberpedia/de/articles/ueberwachtes-unueberwachtes-bestaerkendes-lernen/
  es: https://startbrain.io/cyberpedia/es/articles/aprendizaje-supervisado-no-supervisado-por-refuerzo/
  it: https://startbrain.io/cyberpedia/it/articles/apprendimento-supervisionato-non-supervisionato-rinforzo/
  nl: https://startbrain.io/cyberpedia/nl/articles/gesuperviseerd-niet-gesuperviseerd-bekrachtigingsleren/
  pl: https://startbrain.io/cyberpedia/pl/articles/uczenie-nadzorowane-nienadzorowane-przez-wzmacnianie/
  pt: https://startbrain.io/cyberpedia/pt/articles/aprendizagem-supervisionada-nao-supervisionada-reforco/
  ja: https://startbrain.io/cyberpedia/ja/articles/supervised-unsupervised-reinforcement-learning/
sources:
  - "n": 1
    title: What is Machine Learning? Types and uses
    publisher: Google
    url: https://cloud.google.com/learn/what-is-machine-learning
    accessed: "2026-11-14"
  - "n": 2
    title: "Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations"
    publisher: NIST
    url: https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-2e2025.pdf
    accessed: "2026-11-14"
  - "n": 3
    title: Principles for security of Machine learning ML
    publisher: NCSC (UK)
    url: https://www.ncsc.gov.uk/collection/machine-learning-principles
    published: "2024-05-22"
    accessed: "2026-11-14"
  - "n": 4
    title: Securing Machine Learning Algorithms | ENISA
    publisher: ENISA
    url: https://www.enisa.europa.eu/publications/securing-machine-learning-algorithms
    published: "2021-12-14"
    accessed: "2026-11-14"
  - "n": 5
    title: Deep Learning vs. Machine Learning - Azure Machine Learning
    publisher: Microsoft
    url: https://learn.microsoft.com/en-us/azure/machine-learning/concept-deep-learning-vs-machine-learning?view=azureml-api-2
    accessed: "2026-11-14"
---

# Supervised, unsupervised and reinforcement learning

A typical AI system was not programmed rule by rule. According to Google [1], machine learning lets a system learn and improve on its own, without being explicitly programmed, when it is given large amounts of data. How the data is supplied changes what the system learns.

Google [1] presents four types of models. Three are described here: supervised learning, unsupervised learning and reinforcement learning. The fourth, called semi-supervised, is a mixed approach in which only part of the data is labeled [1]. For an employee, one point matters: a model depends on the data that trained it [1], and an attacker may try to distort that data [2].

## In 30 seconds

- Where it happens: in tools that sort messages, spot suspicious activity or drive an agent, meaning a program that learns by trial and receives feedback on its actions [1][2].
- What the attacker wants: to distort training by inserting or modifying examples, which NIST calls data poisoning [2].
- The reflex: ask where the data used to train a tool came from before trusting its output, because a model is often only as good as the data it is given [1].

## Straight answers

### What is the difference between the three ways of learning?

It comes down to what the system receives during training. In supervised learning, it receives labeled examples [1]. In unsupervised learning, it receives data without labels [1]. In reinforcement learning, an agent interacts with an environment and learns a policy, that is, a way of acting, that maximizes its reward [2].

### Does an AI tool keep learning once it is in use?

The NCSC (UK) [3] defines continual learning as the set of methods that adjust or extend a model's training once it is deployed, based on interactions and usage feedback. Google [1] states for its part that a model can be retrained with new data and receive feedback from real users. NIST [2] notes that detectors of attacks on industrial systems are often retrained with data collected during operation, which gives an attacker the opportunity to poison them.

### What is data poisoning?

NIST [2] defines poisoning attacks as attacks carried out during the training phase. In data poisoning, the adversary controls part of the training data by inserting or modifying examples [2]. ENISA [4] lists data poisoning among the threats aimed at machine learning systems, along with adversarial attacks and data exfiltration.

## How it works

### Supervised learning: learning from corrected examples

The model receives labeled data and learns to link a feature to a label [1]. Google [1] gives the example of recognizing apples: to train the algorithm, it is shown images labeled as apples. The expected answer is known in advance [1].

NIST [2] distinguishes two families of supervised techniques: classification, where the predicted labels are discrete, and regression, where they are continuous. Microsoft [5] notes that the output of a machine learning model is most often a numerical value, such as a score or a classification.

NIST [2] cites spam classifiers trained on example emails. It also describes the case of malware classification, where files may be submitted by attackers to threat intelligence platforms and where labeling is done with antivirus signatures or other external methods [2].

### Unsupervised learning: finding groups without instructions

The model learns patterns from unlabeled data [1]. The "correctness" of the output is not known in advance: the algorithm learns from the data without human intervention and sorts it into groups according to its attributes [1]. If it is given images of apples and bananas, it separates them on its own [1]. Google [1] considers it effective for descriptive modeling and pattern finding, and cites K-means clustering among the common algorithms.

In security, NIST [2] mentions unsupervised models used for malware clustering and for anomaly detection in network traffic. It points out that poisoning has been demonstrated in both of these uses [2].

### Reinforcement learning: learning by trying

Google [1] describes it as "learn by doing", through trial and error. An agent learns a task through a feedback loop, until its performance reaches a desirable level. It receives positive reinforcement when it succeeds and negative reinforcement when it fails [1]. In Google's example [1], researchers taught an algorithm to play go: without knowing the rules, it moved pieces at random and eventually beat a human player.

The NCSC (UK) [3] specifies that its security principles do not specifically cover reinforcement learning, because this technique develops differently. They can serve as a basis, but additional security precautions are probably needed [3].

## Pitfalls

### Poor-quality data

According to Google [1], assuming good-quality training data, the more examples the algorithm receives, the more accurate the model becomes. Conversely, an algorithm fed a biased dataset produces biased results [1]. Obtaining and cleaning structured data is an important first step for supervised learning, and it can be difficult when the data is scattered across several places in the organization [1].

### Poisoned data

NIST [2] writes that poisoning can be organized at scale: an adversary with limited financial means could control a fraction of the public datasets used for training. It can degrade the model on all examples, or cause errors on a small number of targeted examples [2]. A simple method, label flipping, consists of supplying training examples whose label is wrong or altered [2].

Defenses are uneven. Attacks that degrade the model on all examples are generally detectable by monitoring the usual performance metrics, such as precision or recall [2]. Targeted attacks, by contrast, are "notoriously challenging to defend against" according to NIST [2], which cites as measures access control on training data, data cleaning and validation, and attestation of their provenance and integrity.

The NCSC (UK) [3] reminds readers that AI and machine learning systems are subject to new vulnerabilities, to be considered in addition to classic threats, and that security must be a requirement throughout the system's life cycle.

### A model can be fooled without touching its training

An evasion attack modifies the examples presented to an already trained model in order to change its prediction [2]. NIST [2] cites a study of a commercial phishing page detector: out of 4,600 samples that the system judged uncertain, the authors identified 100 adversarial examples. The attackers did not use optimization methods there, but simple techniques such as cropping, masking or image blurring [2].

NIST [2] adds that designing models that resist evasion while keeping their accuracy remains an open problem. Some evasions against malware detectors and against a mail protection system have been demonstrated by researchers, but they have not led to attacks in the wild [2].

### Variable data needs

Microsoft [5] states that machine learning can work with small amounts of data, whereas deep learning needs large amounts of training data.

## Sources

1. Google, "What is Machine Learning? Types and uses". https://cloud.google.com/learn/what-is-machine-learning (accessed 14 November 2026)
2. NIST, "Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations". https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-2e2025.pdf (accessed 14 November 2026)
3. NCSC (UK), "Principles for security of Machine learning ML", 22 May 2024. https://www.ncsc.gov.uk/collection/machine-learning-principles (accessed 14 November 2026)
4. ENISA, "Securing Machine Learning Algorithms | ENISA", 14 December 2021. https://www.enisa.europa.eu/publications/securing-machine-learning-algorithms (accessed 14 November 2026)
5. Microsoft, "Deep Learning vs. Machine Learning - Azure Machine Learning". https://learn.microsoft.com/en-us/azure/machine-learning/concept-deep-learning-vs-machine-learning?view=azureml-api-2 (accessed 14 November 2026)

---

Original source: Cyberpedia, the public cybersecurity and AI encyclopedia published by Brain. Canonical version: https://startbrain.io/cyberpedia/en/articles/supervised-unsupervised-reinforcement-learning/
You may reuse this text freely under the CC BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), provided you credit "Brain, Cyberpedia" with a link to the canonical URL above.
