Network Authentication, Authorization, and Accounting (AAA, pronounced "triple-A") is a technology that has been in use since before the days of the Internet as we know it today. Authentication asks the question, "Who or what are you?" Authorization asks, "What are you allowed to do?" And finally, accounting wants to know, "What did you do?" These fundamental security building blocks are being used in expanded ways today. This article, the first in a two-part series, focuses on the overall concepts of AAA, defines the elements involved in AAA communications, and discusses high-level approaches to achieving specific AAA goals. Part two of the article, to be published in a future issue of IPJ, will discuss the protocols involved, specific AAA applications, and considerations for the future of AAA.
AAA, at its core, is all about enabling mobility and dynamic security. Without AAA, a network must be statically configured to control access; IP addresses must be fixed, systems cannot move, and connectivity options should be well defined. Even the earliest days of dialup access broke this static model, thereby requiring AAA. Today, the proliferation of mobile devices, diverse network consumers, and varied network access methods combine to create an environment that places greater demands on AAA.
AAA has a part to play in almost all the ways we access a network today. Emerging technologies such as Network Access Control (NAC) extend AAA even into corporate Ethernet access (historically the "trusted" network that set the benchmark level of security that all other types of access had to match). Today, wireless hotspots need AAA for security, partitioned networks require AAA to enforce segmentation, and remote access of every kind uses AAA to authorize remote users.
It is not clear when the term AAA first gained acceptance, but an examination of academic papers finds "authentication, authorization, and accounting" used as a discrete term (albeit without the AAA acronym) as early as 1983 in an IEEE paper . Though mired in pre-Internet Open Systems Interconnection (OSI)-centric terminology, the ordering of the "A's" is the same as today's usage.
For most network administrators, the genesis of AAA coincided with the development of the Remote Authentication Dial-In User Service (RADIUS) protocol . RADIUS was developed by Livingston Enterprises (now part of Alcatel-Lucent) in the early 1990s, became an Internet standard through the IETF in 1997, and today is the most widely accepted AAA protocol.
Another widely adopted AAA protocol, which predates RADIUS as an RFC by four years, is the Terminal Access Controller Access Control System (TACACS) . Though never an Internet standard, TACACS evolved into XTACACS and then TACACS+, the latter of which is the only version of TACACS in use today.
Before we delve into the details of these protocols, it is important to understand the roles played within a AAA system.
Core Components of AAA
Client: The client is the device attempting to access the network. The client either authenticates itself, or it acts as a proxy to authenticate the user.
Policy Enforcement Point (Authenticator): The Policy Enforcement Point (PEP) is sometimes called the authenticator or dial-in server, VPN concentrator, firewall, gateway General Packet Radio Service (GPRS) support node, Ethernet switch, wireless access point, or an inline security gateway. The PEP is responsible for enforcing the terms of a client's access. This enforcement varies based on the capabilities of the PEP and is discussed later in this article.
Policy Information Point: The Policy Information Point (PIP) is a repository of information to help make the access decision. It could be a database of device IDs, a user directory such as the Lightweight Directory Access Protocol (LDAP), a one-time password (OTP) token server, or any other system that houses data relevant to a device or user access request.
Policy Decision Point (AAA Server): The Policy Decision Point (PDP) is the brain of the AAA decision. It collects the access request from the client through the PEP. It also queries any relevant PIPs to gather the information it needs to make the access decision. The PDP, as its name implies, is the entity that makes the final decision around network access. It also can send specific authorizations back to the PEP that apply settings or constraints to the client's network traffic.
Accounting and Reporting System: Whether on a dedicated system or built as part of a PDP, tracking use of the network with accounting is one of the best features of AAA. With all forms of network access now offering controlled access, the AAA service can tell you who got on the network, from where, and what that person was granted access to.
It is important to note that the preceding categories are logical containers of functions and not necessarily dedicated physical devices. Often elements are combined, such as PEP with PDP, and PDP with PIP.
Example AAA Flow
Now that we have examined the components of a AAA solution, walking through a typical use case will help cement our understanding of the role that each entity plays. Figure 1 shows an example of a client attempting to gain access to the network.
Figure 1: A Client Connects to a AAA-Protected Network
The client attempts to connect to the network, is challenged for identity information, and sends this information to the PEP. In this example, let's assume the client is a laptop with a worker attempting to access an organization's VPN from a remote location. Additionally, we'll assume this is a valid, permitted use of the network.
The PEP sends the collected identity information to the PDP. In some cases (discussed in part two of this article), the PEP cannot see the specific identity information provided but instead relays the information directly to the PDP.
The PDP queries any configured PIPs for information about the client and validates that the credential provided by the client is valid. In this example, the PIP is an LDAP directory.
The PIP returns a success or failure message from the credential validation step and sends additional information about the client to the PDP for evaluation. This information could include the role of the user, the home location for the user, and so on.
The PDP evaluates information learned about the client through the client, PEP, and PIP; the role of the PEP and PIP that serviced the request; and any contextual information (such as time of day) against its configured policies. Based on this information, the PDP makes an authorization decision.
The PDP sends the PEP the authentication result and any authorizations specific to the client. These authorizations trigger specific PEP actions to apply to the client. For example, the authorization data might trigger specific Access Control Lists (ACLs) or IP pool assignments for the client.
The PDP also sends the result of this transaction to the accounting system.
The PEP applies the authorization profile learned from the PDP and sends the "authentication successful" message to the client. The PEP can also be configured to send accounting information on this new connection to the accounting and reporting system.
The client accesses the production network through the PEP.
Elements of Authentication
When performing authentication, numerous elements can be evaluated before a PDP reaches its access decision. At a high level, these elements can be broken down into three categories: the principal itself (the user, device, or service requesting access), the credential the principal submits (shared key, one-time password, digital certificate, or biometric credential), and the contextual information describing the transaction (location, time of day, software state, and so on).
Principal: The principal is the entity requesting authorization. It is generally some combination of user, device, or service. When concerned with a user, the PIP can provide attributes about the user such as role or group affiliations, job title, e-mail address, physical address, and so on. In specific applications, it can include much more granular information. For example, a higher-education facility might be interested in knowing a student's class schedule when servicing the student's authentication request. When the principal is a device, the same thinking applies. The PIP can inform the PDP if the device is a managed asset, what its basic usage parameters are, and so on. User and device authentication can be carried out sequentially for the same transaction, often involving device authentication first and then user authentication. Lastly, a service such as a network management process can authenticate. In this case, the service almost always looks like a user to the AAA infrastructure and is handled accordingly.
Credential: The next element the PDP considers is the credential the user or device submits as proof of identity. There are four main types of credentials: shared key (password), one-time password (OTP), digital certificate, and biometric credential. This section examines each of these types. The first and most widely used form of credential is the shared key, typically a user password. AAA deployments that use shared keys can be subdivided based on the protocol the system uses to verify the password, including the Password Authentication Protocol (PAP) , Challenge Handshake Authentication Protocol (CHAP) , and Microsoft CHAP Extensions (MS-CHAP) Versions 1  and 2 . PAP authentication is a plaintext authentication method that is not recommended for use in security-sensitive environments.
However, many newer protocols provide a secure transport for PAP, making its use in AAA still quite common. Some of these methods are discussed in part two of this article. CHAP improves on the security of PAP by not sending the password in the clear but rather a challenge based on a hash of the password. MS-CHAP is a Microsoft extension to CHAP that tunes things a little bit for Microsoft environments. Version 2 of MS-CHAP addresses security weaknesses in Version 1. MS-CHAPv2 is quite common today in Microsoft environments. CHAP in all its forms is vulnerable to dictionary attacks because even if a hash cannot be decrypted, common passwords can be guessed and those hash values can be computed.
A second, also widely used credential type is the OTP. At login time, users refer to their personal token to get the OTP they will type in. The token is generally provided in hardware or software form. Tokens are designed to generate seemingly random passwords that are synchronized with a token server acting as a PIP. The OTP can be sent in the clear because it is used only once; after a configurable time (for example, 30 seconds) a new password is generated. When an OTP is combined with a Personal Identification Number (PIN), two-factor authentication is achieved because the client needs to have something (the token) and know something (the PIN).
The third type of credential is the digital certificate. Digital certificates can be stored either locally on the client or on some sort of removable device such as a smartcard. A full discussion of asymmetric-key cryptography is outside the scope of this article, but at a high level, certificates work by asserting the identity of their bearer by having the certificate signed by a trusted Certificate Authority (CA). CAs can be external entities such as a government or commercial enterprise or they can be internal to a given organization. The certificate itself can be freely distributed, because the only way it can be validated as belonging to the rightful owner is in combination with the private key. Because they reside on the client, certificates are most often used to authenticate a physical entity rather than an individual. However, smartcards are changing this paradigm by enabling users to take their digital certificate (and private keys) with them, thereby disassociating the certificate from the machine itself. Similar to an OTP without a PIN, a digital certificate or smartcard alone does not provide two-factor authentication. Certificate deployments, particularly smartcards, are addressing this problem by requiring a PIN to unlock access to the credential.
The fourth and least widely deployed type of credential is the biometric credential. Biometrics  ignores something you have and something you know and instead focus on something you are. Fingerprint scanners, iris scanners, and facial recognition are all forms of biometric authentication. Because biometrics is the newest form of credential, it is currently experiencing heightened anticipation among users regarding potential applications—and also scrutiny for potential weaknesses.
Contextual: The last element the PDP typically considers in its authentication decision is the contextual information associated with the AAA request, including the network and physical location of the request, the type of access provided by the PEP, the time of day, and potentially other elements such as network load, security threat level, and so on. A relatively new entrant into this set of contextual information is client device posture, typically discussed under the rubric of Network Access Control (NAC). NAC or posture checks examine the software state of the client before it connects. NAC data allows the PDP to assess the degree of risk posed by the connecting client before granting the client access to the network. For example, if a system is running an out-of-date operating system, has no current security applications running, or otherwise exhibits high-risk behavior, it may not be granted access to the network. NAC will be discussed in more detail in part two of this article.
At its core, authorization means determining what a client is allowed to do on the network. However, the granularity of this authorization is only as good as the sophistication of the PDP and the enforcement capabilities of the PEP. This section examines the authorization options for network AAA, including Layer 2 segmentation, Layer 3 filtering, and Layer 7 entitlements. It closes with an examination of some of the challenges encountered when sending or "provisioning" the authorizations from the PDP to the PEP.
Null Authorization (Authentication Only): Strangely the most common authorization in AAA is no authorization at all. After the authentication event occurs, the client is immediately granted full access to the network. This characteristic is a holdover from the original goal of remote-access AAA: to perform an authentication check that simply determines whether the client should be trusted as if it were connected to the organization's home network. Because these home networks employed no segmentation or filtering within them, it was natural that remote-access techniques such as dialup and VPN would likewise employ neither. Today however, authentication is increasingly being used for all forms of network access, with a goal of providing clients with network rights commensurate with their role in the organization. This latter goal requires a strong authorization foundation through the cooperation of the PDP and PEP.
Layer 2 Segmentation: For wireless access points and Ethernet switches, the most common form of authorization enforcement is Layer 2 segmentation, which works by splitting the network into multiple logical segments, isolating certain classes of client from one another. This process is most typically achieved by deploying Virtual LANs (VLANs), which separate the members of one VLAN from other VLANs in the same Layer 2 network—even though the VLANs traverse the same physical network infrastructure.
VLANs can be used to restrict access to specific resources by working in coordination with VLAN-specific ACLs on Layer 3 devices upstream from the Layer 2 device. For access points, a given wireless Service Set Identifier (SSID) can be associated with a VLAN on the wired side of the access point. Multiprotocol Label Switching (MPLS) is more commonly associated as a WAN transport, but there is nothing to prevent labels for traffic based on AAA. More commonly, the client is associated with a VLAN and the VLAN is associated with an MPLS label further into the infrastructure.
Layer 3 Filtering: Layer 3 filtering authorizes access to resources through ACLs configured on Layer 3 devices (routers, Ethernet switches, security gateways, and so on). These ACLs (which generally encompass Layer 4 of the OSI stack as well) can enforce authorizations to a range of hosts, specific hosts, or services on those hosts. As mentioned earlier, Layer 3 filtering can be combined with Layer 2 segmentation to provide aggregate authorizations for an entire VLAN. This filtering is the most common technique on network infrastructure devices, whereas security gateways tend to apply ACLs to specific clients. Additionally, technologies such as IP Security (IPsec)  provide a Layer 3 filtering capability by allowing only certain types of traffic to travel through the VPN tunnel.
Layer 7 Entitlements: Increasingly, security gateways are able to go beyond Layer 3 and 4 filtering and are starting to become application-aware, meaning that the authorizations handed from the PDP to the PEP can be very granular, focusing on the specific applications that are needed rather than broader filters based on segments or hosts on the network. Because this technology is still relatively new, there are no standards yet to make this interaction work transparently. As a result, most granular application filters are written on the PEP itself in order to allow the PDP to trigger a preexisting profile on the PEP. These sorts of provisioning challenges are discussed further in the next section.
Provisioning Challenges: In AAA parlance, the term "provisioning" refers to communicating a user's session rights and constraints to the PEP so that the PEP can grant and enforce these permissions. One of the most difficult aspects of provisioning access rights on a PEP is communicating the decision of the PDP in a format the PEP can understand. This fact is one of the reasons that many PEPs come with a lightweight PDP. This approach solves the narrow problem for that PEP but creates management challenges when coordinating network AAA across a broader enterprise, because the enterprise AAA policies must be implemented individually on each unique type of PEP on the network. Because RADIUS is the most commonly used network AAA protocol, it is natural to communicate the PDP decision using that protocol. RADIUS attributes such as the "filter-id" allow the PDP to trigger a preexisting filter on the PEP
In addition, many PEP vendors support Vendor Specific Attributes (VSAs) in RADIUS to enable the PDP to speak the language of the PEP more specifically. This process works well but creates a significant amount of work on the PDP to enable it to translate the policy result and correctly communicate it to each type of PEP. Another option soon to be sanctioned by the standards bodies is an extension to RADIUS that enables the sending of standard IP ACLs using RADIUS attributes .
One further option for provisioning is through the Simple Network Management Protocol (SNMP), which is typically used to assign Layer 2 ports to VLANs or to enable or disable interfaces. This process can work, but remember that the version of SNMP typically in deployment is still SNMPv2c, which is User Datagram Protocol (UDP)-based (connectionless) and unencrypted. Therefore, the SNMP traffic is prone to packet loss when links are congested or devices are busy, thereby requiring costly application layer retransmission schemes. It also means the transmissions themselves are vulnerable to inspection or modification. These attributes make SNMP generally a poor choice for security-sensitive tasks. RADIUS also uses UDP, but supports basic retransmission as part of the protocol.
Another provisioning method used today is standard Secure Shell (SSH) Protocol or HTTPS-based configuration. This method manages a device through standard administrative interfaces to set enforcement techniques. Although this method gives the PDP full access to the features of the PEP, it is very difficult to coordinate the dynamic aspects of the client AAA event with the static elements of the running configuration of the PEP. Finally, new protocols are emerging to make provisioning easier. NETCONF  is an Extensible Markup Language (XML)-based protocol designed as a replacement for network management applications connecting to devices over the command-line interface (CLI).
As this section has shown, there are numerous approaches to authorization in AAA. Each PEP has its own capabilities, but the challenge for a diverse network is to consistently authorize clients, regardless of the given PEP they access the network through.
Accounting is an increasingly critical step in the overall AAA process. Regulatory controls are starting to mandate better auditing of network access. The last stage of AAA, accounting simply records which clients accessed the network, what they were granted access to, and when they disconnected from the network. Accounting has always been widely used in the Internet Service Provide (ISP) space because auditing network access is the basis for billing ISP customers. Increasingly, accounting is being used as a way to correlate client attribute information (username, IP address, etc.) with actions and events on the network.
This correlation can make other systems that are not user-aware more intelligent in the security decisions that they make. For example, a network Intrusion Detection System (IDS) can learn a lot about the behavior of a given IP address. However, when that information is correlated with the user assigned to that IP address—and the permissions that user should have—the relevance of the IDS data increases dramatically.
One of the design considerations of accounting systems is that, given the centralized nature of audit and the decentralized nature of access, they are generally out-of-band with the client's normal communications. This makes them excellent resources to refer to when the network administrator wants to know when the client connected and what the client was granted access to. However, their out-of-band nature makes them poor resources for determining what the client actually did while connected to the network. This information can be learned by the network, as mentioned earlier, by coordinating the AAA accounting information with the rest of the network enforcement and monitoring systems.
Summary and Part Two Teaser
This first part of this article introduced AAA and described many of the foundation concepts necessary to gain a sound understanding of the overall system. After defining the elements involved, a sample flow of a AAA event was described. Additionally, the high-level approaches to authentication, authorization, and accounting were discussed. Part two of this article will discuss the protocols used in AAA, including not just RADIUS, Extensible Authentication Protocol (EAP), TACACS+, and Diameter, but many others. Additionally, specific applications of AAA technology will be described, and some conclusions will be drawn as to what the future holds for AAA.
 Lagsford et. al., "OSI Management and Job Transfer Services," Proceedings of the IEEE, Volume 71, No. 12, December 1983.
 Rigney et. al., "Remote Authentication Dial In User Service (RADIUS)," RFC 2865 (Obsoletes RFC 2138, 2058), June 2000.
 Finseth C., "An Access Control Protocol, Sometimes Called TACACS," RFC 1492, July 1993.
SEAN CONVERY is CTO at Identity Engines, a venture-backed startup developing innovative identity management solutions for enterprise networks. Prior to Identity Engines, Sean (CCIE® no. 4232) worked for seven years at Cisco Systems, most recently in the office of the security CTO. Sean is best known as the principal architect of the SAFE Blueprint from Cisco and the author of Network Security Architectures (Cisco Press, 2004). Sean has presented to or consulted with thousands of enterprise customers around the world on designing secure networks. Before Cisco, Sean held various positions in IT and security consulting during his 14 years in networking. E-mail: email@example.com