{
  "id": "multifactor-authentication",
  "title": "Multifactor Authentication",
  "lang": "en",
  "source_lang": "en",
  "summary": "Multifactor authentication verifies identity using distinct categories of evidence, reducing dependence on passwords while introducing requirements for secure enrollment, recovery, and operation.",
  "category": "computer-science",
  "subcategory": "networks-security",
  "type": "concept",
  "aliases": [
    "Multi-factor authentication",
    "MFA",
    "Multiple-factor authentication"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/multifactor-authentication/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/multifactor-authentication/",
      "json": "https://aiwiki.page/zh/multifactor-authentication.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Multifactor authentication (MFA) is a form of authentication that requires evidence from at least two different categories of authentication factors before accepting a claimed identity. The conventional categories are something a person knows, such as a password; something a person possesses, such as a cryptographic security key; and something a person is, such as a fingerprint. Two-factor authentication (2FA) is the most common instance. Entering two passwords does not constitute two-factor authentication, because both belong to the knowledge category.\n\nMFA is a major control in cybersecurity because the compromise of a password need not, by itself, give an attacker access to an account. Its effectiveness nevertheless depends on the methods used and the entire authentication system surrounding them. A password followed by a text-message code, a security key activated by a personal identification number, and a fingerprint-unlocked passkey can all involve multiple factors, but they differ substantially in resistance to interception, impersonation, device compromise, and account recovery attacks.\n\nMFA applies to consumer accounts, enterprise networks, financial services, government systems, and physical access control. Its study combines cryptography, protocol design, human–computer interaction, and operational security. Distinguishing factor categories from interface steps, and distinguishing authentication from authorization or identity proofing, is essential to understanding what an MFA deployment actually establishes."
    },
    {
      "heading": "Factors, authenticators, and security claims",
      "text": "An authentication factor is a category of evidence, whereas an authenticator is the mechanism or object that produces that evidence. A password is a memorized secret. A hardware security key is a possession authenticator. A biometric sensor measures a physical or behavioral characteristic. One authenticator can incorporate more than one factor: a cryptographic device that will use its private key only after verifying a PIN combines possession with knowledge. Consequently, multifactor authentication need not require separate devices or separate interactions with a remote service.\n\nKnowledge factors include passwords and PINs, but security questions are also knowledge-based and do not add an independent category to a password. Moreover, answers such as a birthplace or a relative’s name may be discoverable rather than secret. Possession factors establish control over something: a device containing a private key, a token sharing a secret with a verifier, or a communication endpoint that receives a challenge. Inherence factors include fingerprints, facial characteristics, and iris patterns. Behavioral measurements, such as typing rhythm, may contribute evidence, although their variability often makes them more suitable for risk assessment than for a decisive authentication check.\n\nThe factors must be meaningfully independent under the system’s threat model. Keeping a password and a one-time-code generator on the same smartphone does not automatically eliminate MFA: compromising one application may not expose the other. However, a fully compromised device might reveal both. Conversely, using two separate devices does not guarantee independence if both can be reset through the same compromised email account. Independence is therefore a property of the implementation and its failure paths, not merely a count of devices.\n\nA simplified model illustrates the potential benefit. If an attacker’s probabilities of defeating two factors are \\(p_1\\) and \\(p_2\\), and those events are independent, the probability of defeating both is\n\n\\[P(A_1 \\cap A_2)=p_1p_2.\\]\n\nActual attacks rarely satisfy complete independence. More generally, the joint probability is \\(P(A_1)P(A_2\\mid A_1)\\). A phishing site that collects both a password and a current verification code creates correlated failures, as does malware controlling the device on which both are entered.\n\nAuthentication establishes confidence that a claimant controls the required authenticators; it does not necessarily establish a civil identity. Identity proofing associates an account with a real person, while authorization determines what an authenticated account may do. A strongly authenticated account can still have excessive permissions, belong to an impostor enrolled under false information, or perform an unauthorized transaction through a compromised application."
    },
    {
      "heading": "One-time passwords and delivered verification codes",
      "text": "A one-time password (OTP) is intended to be accepted only once, or within a tightly limited authentication context. Unlike a reusable password, a captured OTP should not remain useful indefinitely. Common implementations use a shared secret stored by both an authenticator and a verifier, together with a moving value such as an event counter or the current time. The output is shortened to a code that a person can type, commonly six decimal digits.\n\nThe HMAC-based One-Time Password algorithm, HOTP, was specified in RFC 4226 in 2005. It applies a keyed message authentication code to a counter and uses dynamic truncation to derive a decimal output. In simplified notation,\n\n\\[\\operatorname{HOTP}(K,C) =\\operatorname{Truncate}(\\operatorname{HMAC\\!-\\!SHA\\!-\\!1}(K,C)) \\bmod 10^d,\\]\n\nwhere \\(K\\) is the shared secret, \\(C\\) is the counter, and \\(d\\) is the number of digits. The verifier may accept a limited range of future counter values to accommodate button presses that did not result in successful logins. Larger acceptance windows increase both usability and the number of candidate values an attacker can try.\n\nThe Time-Based One-Time Password algorithm, TOTP, was specified in RFC 6238 in 2011. It substitutes a time-derived counter,\n\n\\[T=\\left\\lfloor\\frac{t-T_0}{X}\\right\\rfloor,\\]\n\nwhere \\(t\\) is the current time, \\(T_0\\) is an agreed starting time, and \\(X\\) is the time-step duration. Thirty-second steps are common. Servers frequently tolerate adjacent steps to accommodate clock differences and entry delays, while tracking successful use to prevent replay within an accepted interval. For an ideal six-digit code, one random guess has probability \\(10^{-6}\\) of matching a particular valid value; throttling is therefore essential because short codes cannot provide security against unrestricted guessing.\n\nText-message and voice codes use a different delivery model. A server generates a challenge and sends it through a telephone network. This can provide useful protection against password-only compromise, but possession of a telephone number is not equivalent to possession of a protected cryptographic key. SIM-swap fraud, number reassignment, compromised carrier accounts, and weaknesses in Signalling System No. 7 can expose messages. Email-delivered codes likewise depend on the security and independence of the email account.\n\nOTP systems also create server-side obligations. Shared seeds must be protected, enrollment QR codes must not leak, and code attempts need expiration and rate limits. A stolen seed permits an attacker to generate future codes. Most importantly, manually entered OTPs are not inherently phishing-resistant: an impostor can relay a fresh code to the legitimate service before it expires. “One-time” limits reuse, not real-time interception."
    },
    {
      "heading": "Public-key authenticators, security keys, and passkeys",
      "text": "Public-key authentication replaces a shared authentication secret with a key pair. The authenticator retains a private key, while the service stores the corresponding public key. During login, the service supplies a fresh challenge and the authenticator signs data containing that challenge. The server verifies the resulting digital signature. A stolen public-key database does not, by itself, provide the material needed to impersonate users, unlike a database of shared OTP seeds.\n\nThe FIDO family of protocols applies this approach to interoperable authenticators. Universal 2nd Factor, or U2F, primarily provided a cryptographic second factor alongside a password. FIDO2 combines Web Authentication, the browser-facing standard, with the Client to Authenticator Protocol, CTAP. Authenticators may be external devices connected through USB, NFC, or Bluetooth, or platform authenticators integrated into a phone or computer. A web browser mediates requests between the website and authenticator rather than requiring the user to copy a secret into a webpage.\n\nA central property is binding the credential to the legitimate service. WebAuthn uses a relying-party identifier associated with a domain, and signed authentication data is checked together with client data containing the challenge and origin. A lookalike website cannot ordinarily request a credential for an unrelated legitimate domain. This protection differs fundamentally from a one-time code, which a person can type into any page. Secure contexts and correct origin validation remain necessary; Transport Layer Security protects the channel but does not, on its own, prevent users from connecting securely to an impostor’s domain.\n\nA passkey is a FIDO credential designed for passwordless sign-in. Some passkeys remain bound to a particular device, while others are synchronized through a credential provider. A passkey can support multifactor authentication when use of the private key requires local user verification, such as a device PIN or biometric comparison. The remote service receives a cryptographic assertion and a user-verification indication, rather than the fingerprint image or the PIN. Merely touching a security key may demonstrate user presence without establishing a second factor; presence and verification are distinct protocol concepts.\n\nDevice-bound and synchronized credentials have different operational properties. A non-exportable key in protected hardware limits copying but can complicate replacement after loss. Synchronization improves availability across devices, while making the credential provider’s account protection, device enrollment, and recovery processes relevant to the security analysis. Authenticator attestation can supply information about authenticator provenance or properties, but its usefulness depends on verification policy and must be weighed against privacy and interoperability considerations."
    },
    {
      "heading": "Biometrics and local user verification",
      "text": "Biometrics uses measurable human characteristics to recognize or verify a person. Fingerprint systems extract features such as ridge endings and bifurcations; facial systems derive representations from images; iris systems analyze patterns in the iris. Enrollment creates a reference template, and later samples are compared against it. The process is usually probabilistic rather than an exact comparison, because lighting, pose, sensor noise, aging, and the condition of a finger or face alter measurements.\n\nBiometric systems therefore operate at a decision threshold. A permissive threshold can reduce false rejections of legitimate users while increasing false acceptances of impostors; a stricter threshold generally changes the balance in the opposite direction. Performance figures are meaningful only with their test conditions, population, sensor, and attack assumptions. A false-match rate measured on ordinary comparison samples is not the same as resistance to a deliberately constructed presentation attack, such as an artificial fingerprint or manipulated facial image.\n\nUnlike a password, a bodily characteristic cannot readily be replaced after exposure. Biometric templates also require privacy protections because the underlying characteristics may support identification across contexts. Encryption, access controls, local processing, and limits on retention reduce some risks, but they do not make biometric information revocable in the same sense as a cryptographic key. Presentation-attack detection can inspect depth, texture, motion, or other sensor signals, although its effectiveness depends on the particular implementation and adversary.\n\nA common MFA design avoids sending biometrics to the remote verifier. A device compares the sample locally and, if the comparison succeeds, authorizes a protected private key to sign the authentication challenge. Here the possession factor is control of the device-bound key, and the biometric is an activation factor. The remote service verifies the cryptographic result. This architecture separates biometric matching from online account verification and can avoid a centralized repository of biometric templates.\n\nFallback behavior is part of the design, not an incidental convenience. Many devices permit a PIN or password when biometric verification is unavailable, after reboot, or after repeated unsuccessful attempts. The resulting assurance depends on the strength and rate limiting of that fallback, along with whether hardware prevents unlimited offline guessing. Accessibility also matters: injuries, disabilities, protective equipment, or environmental conditions may prevent reliable use of a particular sensor.\n\nBiometrics establishes neither informed consent nor transaction intent by itself. A successful face or fingerprint check can show that a local matching process accepted a sample, but it does not prove that the person understood the website, payment, or administrative operation being authorized. Systems that need evidence about a specific action require an additional, clearly defined transaction-confirmation mechanism."
    },
    {
      "heading": "Phishing resistance and other attack boundaries",
      "text": "MFA addresses several common consequences of password compromise. Credential stuffing uses username–password pairs obtained elsewhere, exploiting password reuse. Password spraying tests a few common passwords across many accounts. If the attacker lacks a required independent authenticator, successful password guessing or reuse is insufficient for login. However, different MFA methods address different attacks, so counting factors alone does not establish their strength.\n\nPhishing exposes an important distinction. A fraudulent site can request a password and an OTP, then forward both to the real service immediately. An adversary-in-the-middle proxy may relay the entire login interaction and capture the authenticated session. Short code lifetimes do not stop this when the relay operates within seconds. Domain-bound public-key authentication is designed to resist this class of impersonation by preventing an authenticator from producing the required assertion for the wrong relying party.\n\nPush approval introduces a different attack surface. If entering a stolen password triggers a notification on the legitimate user’s phone, an attacker may repeatedly initiate logins until the user approves one through confusion or fatigue. Number matching, in which the user enters or selects a value displayed by the initiating session, can reduce indiscriminate approvals. Displaying the application and contextual details can also help users distinguish requests, but these measures do not necessarily provide cryptographic origin binding or complete resistance to social engineering.\n\nSuccessful MFA usually results in a session credential, often carried in an HTTP cookie. Theft of that credential can allow an attacker to act without repeating authentication. Malware, malicious browser extensions, cross-site scripting, or an exposed application endpoint can therefore undermine an account after a valid login. Session expiration, reauthentication for sensitive actions, token protection, and endpoint security address different portions of this problem. MFA is a control at an authentication boundary, not a replacement for all subsequent security controls.\n\nA compromised verifier also changes the picture. An attacker controlling application logic may bypass factor checks, alter enrollment records, or issue session tokens directly. Public-key authentication reduces the consequences of public-key database disclosure, but cannot guarantee correct behavior by a malicious service. Likewise, an authenticator cannot prevent abuse of permissions legitimately granted to a user, or the manipulation of data after an authorized session has been established.\n\nRecovery and enrollment attacks often avoid the strongest authenticator entirely. An attacker may impersonate a user to a help desk, persuade staff to replace a security key, or exploit a weak password-reset path. A complete assessment consequently examines the easiest path to account control, including legacy protocols and emergency access, rather than evaluating only the normal sign-in screen."
    },
    {
      "heading": "Enrollment, recovery, and authenticator lifecycle",
      "text": "MFA begins before the first protected login. Enrollment binds an authenticator to an account, and that binding must be authorized by evidence appropriate to the account’s risk. Adding an authenticator solely because someone knows the current password can create a takeover window when that password is already stolen. Existing authenticated sessions, previously enrolled factors, identity proofing, administrative approval, or controlled organizational issuance may be used to establish a stronger binding.\n\nEach authenticator type has a provisioning process. A TOTP application commonly receives a shared seed through a QR code, followed by a test code that demonstrates successful configuration. A public-key authenticator generates a key pair and returns registration data; the service stores the credential identifier and public key. A telephone-based factor requires control of a number at enrollment, but continued number ownership cannot simply be assumed indefinitely. Systems must distinguish confirmation of setup from proof of durable, exclusive control.\n\nLoss and replacement are unavoidable operational events. Phones break, security keys disappear, employees change roles, and communication accounts become inaccessible. Recovery codes provide pre-generated secrets that can restore access, but they are typically knowledge or bearer secrets rather than equivalent replacements for a phishing-resistant authenticator. Their storage, entropy, one-time use, and exposure through backups affect their value. Enrolling multiple authenticators can improve availability while increasing the number of credentials whose lifecycle must be managed.\n\nA recovery procedure is effectively an alternative authentication protocol. If a help desk can disable MFA after receiving easily discovered personal details, that procedure may determine the account’s practical security. Stronger recovery can involve previously registered authenticators, independently verified organizational processes, or renewed identity proofing. Delays and notifications can make unauthorized changes more visible, although neither establishes identity by itself. Highly privileged accounts often require recovery controls distinct from those used for ordinary consumer access.\n\nRevocation must address both credentials and active sessions. Removing a lost security key prevents future acceptance of that key, but does not necessarily invalidate sessions established earlier. Resetting a password may not revoke refresh tokens or existing application sessions. Credential removal, session termination, and the invalidation of recovery material are separate operations that must be coordinated when compromise is suspected.\n\nLifecycle records also support accountability. Enrollment, factor replacement, failed challenges, recovery attempts, and administrative overrides can be logged without recording passwords, OTP seeds, biometric samples, or complete recovery codes. Notifications of new authenticators or recovery changes may reveal unauthorized activity, but their usefulness depends on whether the attacker also controls the notification channel. Offboarding similarly requires removal of access across dependent systems, not merely collection of a physical token."
    },
    {
      "heading": "Enterprise architecture, federation, and adaptive authentication",
      "text": "Organizations often centralize authentication in an identity provider rather than implementing MFA independently in every application. The provider authenticates users and communicates results to relying applications through federation protocols such as Security Assertion Markup Language or OpenID Connect. This arrangement can improve policy consistency and reduce repeated prompts, while concentrating security responsibility in the provider, its signing keys, configuration, and administrative accounts.\n\nSingle sign-on and MFA solve different problems. Single sign-on allows one authenticated session to support access to multiple services; MFA determines the evidence required to establish or elevate that session. A user may complete MFA at the identity provider and then access several applications without another challenge. Each application must still determine whether the received assertion meets its requirements, including its intended audience, issuer, validity period, authentication context, and the freshness of authentication.\n\nStep-up authentication requires additional evidence when an action or resource demands greater assurance than the current session provides. Viewing a general internal page and changing a payroll destination need not have identical requirements. A step-up flow might request a phishing-resistant authenticator or require recent user verification. The application must bind the completed step-up to the correct user and session; a generic indication that MFA happened at some unspecified time may be inadequate.\n\nRisk-adaptive authentication varies requirements using signals such as device management status, network location, unusual access patterns, or indicators of automated attack. Anomaly detection and machine learning can support this assessment, but these signals are not automatically independent authentication factors. An IP address may be shared by many people, a browser fingerprint can change or be imitated, and travel estimates can be distorted by network routing or virtual private networks. Risk scoring influences when authentication is required rather than necessarily proving who is present.\n\nCoverage is an architectural issue. An organization may protect its browser login while leaving an older mail protocol, remote-access gateway, administrative interface, or application programming interface accessible through weaker credentials. App-specific passwords sometimes bridge systems that cannot perform interactive MFA, but their existence creates a separate credential class requiring explicit controls. A visible MFA prompt is therefore not evidence that every account-access path is protected.\n\nHuman MFA also differs from machine authentication. A scheduled service, workload, or automated deployment cannot reasonably approve phone notifications. Workload identities, short-lived credentials, mutually authenticated channels, and protected cryptographic keys address those environments. Secrets may be guarded by a hardware security module, but that does not make every automated cryptographic exchange multifactor authentication. The relevant claim depends on what independent evidence is actually required and verified."
    },
    {
      "heading": "Usability, accessibility, and operational measurement",
      "text": "Authentication security depends partly on whether legitimate users can complete the intended process reliably. A mechanism with strong cryptographic properties may fail operationally if it causes frequent lockouts, encourages insecure workarounds, or makes staff routinely override controls. Conversely, a low-friction method can be weak if users cannot distinguish legitimate requests from attacker-generated ones. Usability and security therefore interact through concrete behavior rather than forming a simple opposition.\n\nDifferent methods impose different burdens. Copying a six-digit code requires switching attention and sometimes switching devices. A hardware key may be quick to use but require the correct connector or radio capability. Text messages depend on network delivery and telephone access. Push notifications require a functioning enrolled device and can make unsolicited approval requests seem routine. Passkeys can simplify normal sign-in, while device changes and cross-platform recovery introduce their own sources of confusion.\n\nAccessibility requirements include vision, hearing, dexterity, cognition, and access to assistive technology. A system that relies exclusively on recognizing an image, reading a short-lived code, or physically touching a small sensor can exclude some users. Workplace conditions also matter: gloves, shared workstations, restricted radio environments, and areas where phones are prohibited alter the practical suitability of authenticators. Providing alternative methods need not mean making every method available to every account; alternatives can be selected according to both access needs and assurance requirements.\n\nEconomic and privacy constraints also influence deployment. Requiring employees or customers to own a particular smartphone can transfer costs and expose personal information. Telephone numbers may reveal more identity information than is necessary for cryptographic authentication. Central biometric processing has different privacy consequences from local biometric activation of a key. Logs containing device attributes and location estimates can support fraud analysis while also becoming sensitive records subject to access and retention controls.\n\nOperational measurement needs denominators and context. Useful measures include enrollment completion, login success, challenge abandonment, recovery frequency, help-desk workload, and confirmed account compromise by authentication method. Raw counts can mislead: a method used by a small high-risk population cannot be compared directly with one used by millions of ordinary accounts. Reported reductions in compromise also depend on attack mix, enforcement coverage, and whether bypass and recovery incidents are included.\n\nTesting extends beyond a successful demonstration login. It includes expired challenges, clock drift, duplicate submissions, lost devices, interrupted registration, accessibility tools, and inconsistent federation claims. Security testing examines whether alternate interfaces can bypass MFA or attach an attacker-controlled factor. Because availability is itself a security concern, systems also need defined behavior during identity-provider or messaging outages. Automatically disabling MFA during an outage changes the security model precisely when operational pressure is greatest."
    },
    {
      "heading": "Historical development and standards",
      "text": "Multifactor authentication predates modern internet services. A bank card used with a PIN combines a possession object with a memorized secret, although the strength of that combination depends on how the card and PIN are checked and protected. Enterprise remote-access systems later made extensive use of dedicated tokens displaying changing codes. As online services expanded, software tokens and telephone-delivered challenges brought similar ideas to a much larger population.\n\nStandardized OTP algorithms improved interoperability. The Initiative for Open Authentication, commonly known as OATH, developed specifications for mechanisms such as HOTP and TOTP; it is distinct from OAuth, an authorization framework. Publication of HOTP in 2005 and TOTP in 2011 through the Internet Engineering Task Force allowed multiple vendors to implement compatible code generators and verifiers. The same interoperability also made secure handling of exportable shared seeds an important operational concern.\n\nThe FIDO Alliance, formed in 2012, pursued authentication based on public-key credentials and interoperable authenticators. U2F emphasized a strong second factor usable with existing passwords, while the later FIDO2 architecture supported both second-factor and passwordless operation. The World Wide Web Consortium published Web Authentication Level 1 as a Recommendation in March 2019. Its integration into browsers and operating systems made cryptographic authentication available through standard web interfaces rather than proprietary browser extensions.\n\nPublic-sector guidance helped distinguish authenticator strength from factor count. The United States National Institute of Standards and Technology’s Digital Identity Guidelines, including the SP 800-63B family, describe authenticator requirements and authentication assurance levels. These frameworks distinguish matters such as replay resistance, phishing resistance, authenticator activation, and protected key storage. A system’s assurance level cannot be inferred merely from a vendor calling its login flow “MFA.”\n\nPayment regulation provides a separate example of the same conceptual structure. The European Union’s revised Payment Services Directive and associated regulatory technical standards established strong customer authentication requirements based on independent elements of knowledge, possession, and inherence, subject to the framework’s scope and exemptions. For relevant remote electronic payments, dynamic linking connects authentication to the payment amount and payee. This adds transaction-specific protection beyond simply proving access to an account.\n\nStandards terminology continues to matter as passwordless authentication becomes more common. “Passwordless” means that a password is not used in the particular flow; it does not specify whether one factor or several are present. Similarly, “two-step verification” describes an interaction sequence without necessarily establishing different factor categories. Evaluating a modern implementation requires identifying the actual authenticator, its activation requirements, the verifier’s checks, and the recovery paths that remain available."
    }
  ],
  "links": [
    "authentication",
    "cybersecurity",
    "cryptography",
    "smartphone",
    "authorization",
    "digital-signature",
    "web-browser",
    "transport-layer-security",
    "encryption",
    "phishing",
    "social-engineering",
    "http-cookie",
    "malware",
    "identity-provider",
    "anomaly-detection",
    "machine-learning",
    "application-programming-interface",
    "internet-engineering-task-force",
    "operating-system"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}