Transport Layer Security

The Transport Layer Security family of protocols were initially proposed under the name Secure Socket Layer (SSL). The first deployments used this name and many researchers still refer to this security protocol as SSL [FKC1996]. In this chapter, we use the official name that was standardized by the IETF: TLS for Transport Layer Security.

The TLS protocol was designed to be usable by a wide range of applications that use the transport layer to reliably exchange information. TLS is mainly used over the TCP protocol. There are variants of TLS that operate over SCTP RFC 3436 or UDP RFC 6347, but these are outside the scope of this chapter.

A TLS session operates over a TCP connection. TLS is responsible for the encryption and the authentication of the SDUs exchanged by the application layer protocol while TCP provides the reliable delivery of this encrypted and authenticated bytestream. TLS is used by many different application layer protocols. The most frequent ones are HTTP (HTTP over TLS is called HTTPS), SMTP RFC 3207 or POP and IMAP RFC 2595, but proprietary application-layer protocols also use TLS [AM2019].

A TLS session can be initiated in two different ways. First, the application can use a dedicated TCP port number for application layer protocol x-over-TLS. This is the solution used by many HTTP servers that reserve port \(443\) for HTTP over TLS. This solution works, but it requires to reserve two ports for each application : one where the application-layer protocol is used directly over TCP and another one where the application-layer protocol is used over TLS. Given the limited number of TCP ports that are available, this is not a scalable solution. The table below provides some of the reserved port numbers for application layer protocols on top of TLS.

Application

TCP port

TLS port

POP3

110

995

IMAP

143

993

NNTP

119

563

HTTP

80

443

FTP

21

990

A second approach to initiate a TLS session is to use the standard TCP port number for the application layer protocol and define a special message in this protocol to trigger the start of the TLS session. This is the solution used for SMTP with the STARTTLS message. This extension to SMTP RFC 3207 defines the new STARTTLS command. The client can issue this command to indicate to the server that it wants to start a TLS session as shown in the example below captured during a session on port 25.

220 server.example.org ESMTP
EHLO client.example.net
250-server.example.org
250-PIPELINING
250-SIZE 250000000
250-ETRN
250-STARTTLS
250-ENHANCEDSTATUSCODES
250-8BITMIME
250 DSN
STARTTLS
220 2.0.0 Ready to start TLS

In the remaining parts of this chapter, we assume that the TLS session starts immediately after the establishment of the TCP connection. This corresponds to the deployments on web servers. We focus our presentation of TLS on this very popular use case. TLS is a complex protocol that supports other features than the one used by web servers. A more detailed presentation of TLS may be found in [KPS2002] and [Ristic2015].

A TLS session is divided in two phases: the handshake and the data transfer. During the handshake, the client and the server negotiate the security parameters and the keys that will be used to secure the data transfer. During the second phase, all the messages exchanged are encrypted and authenticated with the negotiated algorithms and keys.

The TLS handshake

When used to interact with a regular web server, the TLS handshake has three important objectives:

  1. Securely negotiate the cryptographic algorithms that will be used by the client and the server on the TLS session

  2. Verify that the client interacts with a valid server

  3. Securely agree on the keys that will be used to encrypt and authenticate the messages exchanged over the TLS session

The TLS handshake is a four-way handshake illustrated in the figure below.

msc {
a [label="", linecolour=white],
b [label="Client",linecolour=black],
z [label="", linecolour=white],
c [label="Server", linecolour=black],
d [label="", linecolour=white];

b>>c [ label = "ClientHello[Random]", arcskip="1"];

c>>b [ label = "ServerHello[Random], Certificate", arcskip="1"];
|||;
b>>c [ label = "E(K,MasterSecret), Finished=MAC(MasterSecret||Handshake)", arcskip="1"];
|||;
c>>b [ label = "Finished=MAC(MasterSecret||Handshake)", arcskip="1"];
|||;
c>>b [ label = "Encrypted Record", linecolour="red", textcolour="red"];
b>>c [ label = "Encrypted Record", linecolour="red", textcolour="red"];
}

In a nutshell, the client starts the TLS handshake by proposing a random nonce. The server replies with its random nonce and a certificate that binds its name to a public key. The client generates a MasterSecret that will be used later to derive the session keys and encrypts it with the public key of the server. It also generates a Finished message that contains a MAC of all the messages exchanged to allow the server to detect any modification of the messages sent by the client. The server also sends its own Finished message. At that point, the client and the server sent encrypted records thanks to the keys derived from the MasterSecret.

Let us first discuss the negotiation of the cryptographic algorithms and parameters. Like all security protocols, TLS includes some agility in its design since new cryptographic algorithms appear over the years and some older algorithms become deprecated once cryptanalysts find flaws. The TLS handshakes starts with the ClientHello message that is sent by the client. This message carries the following information :

  • Protocol version number: this is the version of the TLS protocol supported by the client. The server should use the same version of the TLS protocol as the client, but may opt for an older version. Both versions 1.2 and 1.3 of TLS are deployed today. Older versions are being deprecated.

  • Random number: security protocols rely on random numbers. The client sends a 32 bytes long random number where usually four of these bytes correspond to the client’s clock. This random number is used, together with the server’s random number, as a seed to generate the security keys.

  • Cipher suites : this ordered list contains the set of cryptographic algorithms that are supported by the client, with the most preferred one listed first. In contrast with ssh that allows to negotiate independent algorithms for encryption, key exchange and authentication, TLS relies on suites that combine these algorithms together. Many cryptographic suites have been defined for TLS. Various recommendations have been published on the security of some of these suites RFC 7525.

  • Compression algorithm : the client may propose the utilization of a specific compression algorithm (e.g. zlib). In theory, compressing the data before encrypting it is an intelligent way to reduce the amount of data exchanged. Unfortunately, its implementation in TLS has caused several security problems [PHG2013]. For this reason, compression is usually disabled in TLS RFC 7525.

  • Extensions : TLS supports various extensions in the ClientHello message. These extensions RFC 6066 are important to allow the protocol to evolve, but many of them go beyond the scope of this chapter.

Note

The Server Name Indication (SNI)

The Server Name Indication (SNI) extension defined in RFC 6066 is an important TLS extension for web servers. It is used by the client to indicate the name of the server that it wishes to contact. The IP address associated to this name has been queried from the DNS and used to establish the TCP connection. Why should the client indicate the server name in the TLS ClientHello ? The motivation is the same as for the Host header line in HTTP/1.0. With the SNI extension, a single TLS server can support several web sites that use different domain names. Thanks to the SNI extension, the server knows the concerned domain name at the start of the TLS session. Without this extension, hosting providers would have been forced use one IP address per TLS-enabled server.

The server replies to the ClientHello with several messages:

  • the ServerHello message that contains the protocol version chosen by the server (assumed to be the same as the client version in this chapter), the 32 random bytes chosen by the server, the Cipher Suite selected by the server from the list advertised by the client and a Session Id. This Session Id is an identifier which is chosen by the server. It identifies the TLS session and the security parameters (algorithms and keys) negotiated for this session. It is used to support session resumption.

  • the Certificate message provides the certificate (or usually a chain of certificates) that binds a domain name to the public key used by the server. TLS uses the server certificates to authenticate the server. It relies on a Public Key Infrastructure that is composed of a set of root certification authorities that issue certificates to certification authorities that in the end issue certificates to servers. TLS clients are usually configured with the public keys of several root certification authorities and use this information to validate the certificates that they receive from servers. For historical reasons, the TLS certificates are encoded in ASN.1 format. The details of the ASN.1 syntax [Dubuisson2000] are outside the scope of this book.

  • the ServerKeyExchange message is used by the server to transmit the information that is required to perform the key exchange. The content of this message is function of the selected key exchange algorithm.

  • the ServerHelloDone indicates that the server has sent all the messages for the first phase of the handshake.

At this point, it is time to describe the TLS key exchange. TLS supports different key exchange mechanisms that can be negotiated as part of the selection of the cipher suite. We focus on two of them to highlight their differences:

  • RSA. This key exchange algorithm uses the encryption capabilities of the RSA public-key algorithm. The client has validated the server’s public key thanks to the Certificate message. It then generates a (48 bytes) random number, encrypts it with the server public key and sends the encrypted number to the server in the ClientKeyExchange message. The server uses its private key to decrypt the random number. At this point, the client and the server share the same (48 bytes long) secret and use it to derive the secret keys required to encrypt and authenticate data in the second phase. With this key exchange algorithm, the server does not need to send a ServerKeyExchange message.

  • DHE_RSA. This key exchange algorithm is the Ephemeral Diffie Hellman key exchange with RSA signatures to authenticate the key exchange. It operates as a classical authenticated Diffie Hellman key exchange. If this key exchange has been selected by the server, it sends its Diffie Hellman parameters in the ServerKeyExchange message and signs them with its private key. The client then continues the key exchange and sends the results of its own computation in the ClientKeyExchange message. DHE_RSA is thus an authenticated Diffie Hellman key exchange where the initial message is sent by the server (instead of the client as in our first example but since the protocol is symmetric, this does not matter).

An important difference between DHE_RSA and RSA is their reaction against attacks. DHE_RSA is considered by many to be stronger than RSA because it supports Perfect Forward Secrecy. This property is important against attackers that are able to eavesdrop all the (encrypted) data sent and received by a server. Consider that Terrence is such an attacker that has stored all the packets exchanged by Bob’s server during the last six months. If he manages, by any means, to obtain Bob’s private key, he will be able to decrypt all the keys used to secure the TLS sessions with Bob’s server during this period. With DHE_RSA, a similar attack is less devastating. If Terrence knows Bob’s private key, he will be able to launch a man-in-the-middle attack against future TLS sessions with Bob’s server. However, he will not be able to recover the keys used for all the past sessions that he captured.

Note

Perfect Forward Secrecy

Perfect Forward Secrecy (PFS) is an important property for key exchange protocols. A protocol provides PFS if its design guarantees that the keys used for former sessions will not be compromised even if the private key of the server is compromised. This is a very important property. DHE_RSA provides Perfect Forward Secrecy, but the RSA key exchange does not provide this property. In practice, DHE_RSA is costly from a computational viewpoint. Recent implementations of TLS prefer to thus ECDHE_RSA or ECDHE_ECDSA when Perfect Forward Secrecy is required.

All the information required for the key exchange has now been transmitted. There are two important messages that will be sent by the client and the server to conclude the handshake and start the data transfer phase.

The client sends the ChangeCipherSpec message followed by the Finished message. The ChangeCipherSpec message indicates that the client has received all the information required to generate the security keys for this TLS session. This messages can also appear later in the session to indicate a change in the encryption algorithms that are used, but this usage is outside the scope of this book. The Finished message is more important. It confirms to the server that the TLS handshake has been performed correctly and that no attacker has been able to modify the data sent by the client or the server. This is the first message that is encrypted with the selected security keys. It contains a hash of all the messages that were exchanged during the handshake.

The server also sends a ChangeCipherSpec message followed by a Finished message.

Note

TLS Cipher suites

A TLS cipher suite is usually represented as an ASCII string that starts with TLS and contains the acronym of the key exchange algorithm, the encryption scheme with the key size and its mode of operation and the authentication algorithm. For example, TLS_DHE_RSA_WITH_AES_128_GCM_SHA256 is a TLS cipher suite that uses the DHE_RSA key exchange algorithm with 128 bits AES in GCM mode for encryption and SHA-256 for authentication. The official list of TLS cipher suites is maintained by IANA 1. The NULL acronym indicates that no algorithm has been specified. For example, TLS_ECDH_RSA_WITH_NULL_SHA is a cipher suite that does not use any encryption but still uses the ECDH_RSA key exchange and SHA for authentication.

The TLS record protocol

The handshake is now finished. The client and the server will exchange authenticated and encrypted records. TLS defines different formats for the records depending on the cryptographic algorithms that have been negotiated for the session. A detailed discussion of these different types of records is outside the scope of this introduction. For illustration, we briefly describe one record format.

As other security protocols, TLS uses different keys to encrypt and authenticate records. These keys are derived from the MasterSecret that is either randomly generated by the client after the RSA key exchange or derived from the Diffie Hellman parameters after the DH_RSA key exchange. The exact algorithm used to derive the keys is defined in RFC 5246.

A TLS record is always composed of four different fields :

  • a Type that indicates the type of record. The most frequent type is application data which corresponds to a record containing encrypted data. The other types are handshake, change_cipher_spec and alert.

  • a Protocol Version field that indicates the version of the TLS protocol used. This version is composed of two sub fields : a major and a minor version number.

  • a Length field. A TLS record cannot be longer than 16,384 bytes.

  • a TLSPlainText that contains the encrypted data

TLS supports several methods to encrypted records. The selected method depends on the cryptographic algorithms that have been negotiated for the TLS session. A detailed presentation of the different methods that can be used to produce the TLSPlainText from the user data is outside the scope of this book. As an example, we study one method: Stream Encryption. This method is used with cryptographic algorithms which can operate on a stream of bytes. The method starts with a sequence of bytes provided by the user application: the plain text. The first step is to compute the authentication code to verify the integrity of the data. For this, TLS computes \(MAC(SeqNum, Header, PlainText)\) using HMAC where SeqNum is a sequence number which is incremented by one for each new TLS record transmitted. The Header is the header of the TLS record described above and PlainText is the information that needs to be encrypted. Note that the sequence number is maintained at the two endpoints of the TLS session, but it is not transmitted inside the TLS record. This sequence number is used to prevent replay attacks.

Note

MAC-then-encrypt or Encrypt-then-MAC

When secure protocols use Message Authentication and Encryption, they need to specify how these two algorithms are combined. A first solution, which is used by the current version of TLS, is to compute the authentication code and then encrypt both the data and the authentication code. A drawback of this approach is that the receiver of an encrypted TLS record must first attempt to decrypt data that has potentially been modified by an attacker before being able to verify the authenticity of the record. A better approach is for the sender to first encrypt the data and then compute the authentication code over the encrypted data. This is the encrypt-then-MAC approach proposed in RFC 7366. With encrypt-then-MAC, the receiver first checks the authentication code before attempting to decrypt the record.

Improving TLS

During the last two decades, the deployment of TLS has continued to grow. The early TLS servers were only used for critical services such as e-commerce websites or online banks. As CPU performance improved, it became much more cost-effective to use TLS to secure non-critical parts of web servers, including the delivery of HTML pages and even video services. There is now a growing number of applications that rely on TLS [AM2019].

In 2013, the statistics collected by the Firefox Telemetry project 3 revealed that 30% of the web pages loaded by Firefox users were done over HTTPS. In October 2019, 80% of the web pages are loaded over HTTPS. In six years, HTTPS became the dominant protocol to access web services. Another look at the deployment of HTTPS on web sites may be found in [Helme2019].

Measurement studies that analyzed the evolution of TLS over the years have identified several important changes in the TLS ecosystem [KRA2018]. First, the preferred cryptographic algorithms have changed. While RC4 was used by 60% of the connections in 2012, its usage has dropped since 2015. AEA started to be deployed in 2013 and is now used for more than 90% of the connections. The deployed versions of TLS have also changed. TLS 1.0 and TLS 1.1 are now rarely used. The deployment of TLS 1.2 started in 2013 and reached 70% of the connections in 2015. Version 1.3 of TLS, that is described below is also widely deployed.

Another interesting fact is are the key exchange schemes. In 2012, RSA was the dominant solution, used by more than 80% of the observed connections [KRA2018]. In 2013, Edward Snowden revealed the surveillance activities of several governments. These revelations had a huge impact on the Internet community. The IETF, which standardizes Internet protocols, considered in RFC 7258 that such pervasive monitoring was an attack. Since then, several IETF working groups have developed solutions to counter pervasive monitoring. One of these solutions is to encourage Perfect Forward Security. Within TLS, this implies replacing RSA by an authenticated Diffie Hellman key exchange such as ECDHE. Measurements indicate that since summer 2014, ECDHE is more popular than RSA. In 2018, more than 90% of the observed TLS connections used ECDHE.

The last point is the difficulty of deploying TLS servers [KMS2017]. When TLS servers are installed, the system administrator needs to obtain certificates and configure a range of servers. Initially, getting certificates was complex and costly, but initiatives such as https://letsencrypt.org have simplified this workflow.

In 2014, the IETF TLS working started to work on the development of version 1.3 of the TLS protocol. Their main objectives [Rescorla2015] for this new version were:

  • simplify the design by removing unused or unsafe protocol features

  • improve the security of TLS by leveraging the lessons learned from TLS 1.2 and some documented attacks

  • improve the privacy of the protocol

  • reduce the latency of TLS

Since 2014, latency has become an important concern for web services. As access networks bandwidth continue to grow, latency is becoming a key factor that affects the performance of interactive web services. With TLS 1.2, the download of a web page requires a minimum of four round-trip-times, one to create the underlying TCP connection, one to exchange the ClientHello/ServerHello, one to exchange the keys and then one to send the HTTP GET and retrieve the response. This can be very long when the server is not near the client. TLS 1.3 aimed at reducing this handshake to one round-trip-time and even zero by placing some of the cryptographic handshake in the TCP handshake. This part will be discussed in the TCP chapter. We focus here on the reducing the TLS handshake to a single round-trip-time.

To simplify both the design and the implementations, TLS 1.3 uses only a small number of cipher suites. Five of them are specified in RFC 8446 and TLS_AES_128_GCM_SHA256 must be supported by all implementations. To ensure privacy, all cipher suites that did not provide Perfect Forward Secrecy have been removed. Compression has also been removed from TLS since several attacks on TLS 1.2 exploited its compression capability RFC 7457.

Note

Enterprises, privacy and TLS

By supporting only cipher suites that provide Perfect Forward Secrecy in TLS 1.3, the IETF aims at protecting the privacy of users against a wide range of attacks. However, this choice has resulted in intense debates in some enterprises. Some enterprises, notably in financial organizations, have deployed TLS, but wish to be able to decrypt TLS traffic for various security-related activities. These enterprises tried to lobby within the IETF to maintain RSA-based cipher suites that do not provide Perfect Forward Secrecy. Their arguments did not convince the IETF. Eventually, these enterprises moved to ETSI, another standardization body, and convinced them to adopt entreprise TLS, a variant of TLS 1.3 that does not provide Perfect Forward Secrecy [eTLS2018].

The TLS 1.3 handshake differs from the TLS 1.2 handshake in several ways. First the TLS 1.3 handshake requires a single round-trip-time when the client connects for the first time to a server. To achieve this, the TLS designers look at the TLS 1.2 handshake in details and found that the first round-trip-time is mainly used to select the set of cryptographic algorithms and the cryptographic exchange scheme that will be used over the TLS session. TLS 1.3 drastically simplifies this negotiation by requiring to use the Diffie Hellman can exchange with a small set of possible parameters. This means that the client can guess the parameters used by the server (i.e. the modulus, p and the base g) and immediately start the Diffie Hellman exchange. A simplified version of the TLS 1.3 handshake is shown in the figure below.

msc {
a [label="", linecolour=white],
b [label="Client",linecolour=black],
z [label="", linecolour=white],
c [label="Server", linecolour=black],
d [label="", linecolour=white];

b>>c [ label = "ClientHello[Random, g^c]", arcskip="1"];

c>>b [ label = "ServerHello[Random, g^s]", arcskip="1"];
c>>b [ label = "Certificate, Sign(K,Handshake), Finished, Encrypted Record", textcolour="red", linecolour="red", arcskip="1"];
|||;
b>>c [ label = "Finished", textcolour="red", linecolour="red", arcskip="1"];
c>>b [ label = "Encrypted Record", textcolour="red", linecolour="red", arcskip="1"];
b>>c [ label = "Encrypted Record", textcolour="red", linecolour="red", arcskip="1"];
}

There are several important differences with the TLS 1.2 handshake. First, the Diffie Hellman key exchange is required in TLS 1.3 and this exchange is initiated by the client (before having validated the server identity). To initiate the Diffie Hellman key exchange, the client needs to guess the modulus and the base that can be accepted by the server. Either the client uses standard parameters that most server supports or the client remembers the last modulus/base that it used with this particular server. If the client guessed incorrectly, the server replies with the parameters that it expects and one round-trip-time is lost. When the server sends its ServerHello, it already knows the session key. This implies that the server can encrypt all subsequent messages. After one round-trip-time, all data exchanged over the TLS 1.3 session is encrypted and authenticated. In TLS 1.3, the server certificate is encrypted with the session key, as well as the Finished message. The server signs the handshake to confirm that it owns the public key of its certificate. If the server wants to send application data, it can already encrypt it and send it to the client. Upon reception of the server Certificate, the client verifies it and checks the signature of the handshake and the Finished message. The client confirms the end of the handshake by sending its own Finished message. At that time, the client can send encrypted data. This means that the client only had to wait one round-trip-time before sending encrypted data. This is much faster than with TLS 1.2.

For some applications, waiting one round-trip-time before being able to send data is too long. TLS 1.3 allows the client to send encrypted data immediately after the ClientHello, without having to wait for the ServerHello message. At this point in the handshake, the client cannot know the key that will be derived by the Diffie Hellman key exchange. The trick is that the server and the client need to have previously agreed on a pre-shared-key. This key could be negotiated out of band, but usually it was exchanged over a previous TLS session between the client and the server. Both the client and the server can store this key in their cache. When the client creates a new TLS session to a server, it checks whether it already knows a pre-shared key for this server. If so, the client announces the identifier of this key in its ClientHello message. Thanks to this identifier, the server can recover the key and use it to decrypt the 0-rtt Encrypted record. A simplified version of the 0-rtt TLS 1.3 handshake 2 is shown in the figure below.

msc {
a [label="", linecolour=white],
b [label="Client",linecolour=black],
z [label="", linecolour=white],
c [label="Server", linecolour=black],
d [label="", linecolour=white];

b>>c [ label = "ClientHello[Random, g^c,server_conf=abcd]", arcskip="1"];
b>>c [ label = "0-rtt Encrypted record", textcolour="magenta", arcskip="1"];

c>>b [ label = "ServerHello[Random, g^s]", arcskip="1"];
c>>b [ label = "Certificate, Sign(K,Handshake), Finished, Encrypted Record", textcolour="red", linecolour="red", arcskip="1"];
|||;
b>>c [ label = "Finished", textcolour="red", linecolour="red", arcskip="1"];
c>>b [ label = "Encrypted Record", textcolour="red", linecolour="red", arcskip="1"];
b>>c [ label = "Encrypted Record", textcolour="red", linecolour="red", arcskip="1"];
}

On the web, TLS clients use certificates to authenticate servers but the clients are not authenticated. However, there are environments such as enterprise networks where servers may need to authenticate clients as well. A popular deployment is to authenticate remote clients who wish to access the enterprise network through a Virtual Private Network service. Some of these services run above TLS (or more precisely a variant of TLS named DTLS that runs above UDP [MoR2004] but is outside the scope of this chapter). In such services, each client is authenticated thanks to a public key and a certificate that is trusted by the servers. To establish a TLS session, such a client needs to prove that it owns the public key associated with the certificate. This is done by the server thanks to the CertificateRequest message. The TLS handshake becomes the following one:

msc {
a [label="", linecolour=white],
b [label="Client",linecolour=black],
z [label="", linecolour=white],
c [label="Server", linecolour=black],
d [label="", linecolour=white];

b>>c [ label = "ClientHello[Random, g^c,server_conf=abcd]", arcskip="1"];
b>>c [ label = "0-rtt Encrypted record", textcolour="magenta", arcskip="1"];

c>>b [ label = "ServerHello[Random, g^s]", arcskip="1"];
c>>b [ label = "CertificateRequest, Certificate, Sign(K,Handshake), Finished, Encrypted Record", textcolour="red", linecolour="red", arcskip="1"];
|||;
b>>c [ label = "Certificate, Sign(Kc, Handshake), Finished", textcolour="red", linecolour="red", arcskip="1"];
c>>b [ label = "Encrypted Record", textcolour="red", linecolour="red", arcskip="1"];
b>>c [ label = "Encrypted Record", textcolour="red", linecolour="red", arcskip="1"];
}

The server sends a CertificatRequest message. The client returns its certificate and signs the Handshake with is private key. This confirms to the server that the client owns the public key indicated in its certificate.

There are many more differences between TLS 1.2 and TLS 1.3. Additional details may be found in their respective specifications, RFC 5246 and RFC 8446.

Footnotes

1

See http://www.iana.org/assignments/tls-parameters/tls-parameters.xhtml#tls-parameters-4

2

A detailed explanation of the TLS 1.3 handshake may be found at https://tls13.ulfheim.net/

3

See https://letsencrypt.org/stats/ for a graph and https://docs.telemetry.mozilla.org/datasets/other/ssl/reference.html for additional information on the dataset