Stream cipher

Types

A stream cipher generates successive elements of the keystream based on an internal state. This state is updated in essentially two ways: if the state changes independently of the plaintext or ciphertext messages, the cipher is classified as a synchronous stream cipher. By contrast, self-synchronising stream ciphers update their state based on previous plaintext or ciphertext digits. A system that incorporates the plaintext into the key is also known as an autokey cipher or autoclave cipher.

Synchronous stream ciphers

In a synchronous stream cipher a stream of pseudorandom digits is generated independently of the plaintext and ciphertext messages, and then combined with the plaintext (to encrypt) or the ciphertext (to decrypt). In the most common form, binary digits are used (bits), and the keystream is combined with the plaintext using the exclusive or operation (XOR). This is termed a binary additive stream cipher.

In a synchronous stream cipher, the sender and receiver must be exactly in step for decryption to be successful. If digits are added or removed from the message during transmission, synchronisation is lost. To restore synchronisation, various offsets can be tried systematically to obtain the correct decryption. Another approach is to tag the ciphertext with markers at regular points in the output.

If, however, a digit is corrupted in transmission, rather than added or lost, only a single digit in the plaintext is affected and the error does not propagate to other parts of the message. This property is useful when the transmission error rate is high; however, it makes it less likely the error would be detected without further mechanisms. Moreover, because of this property, synchronous stream ciphers are very susceptible to active attacks: if an attacker can change a digit in the ciphertext, they might be able to make predictable changes to the corresponding plaintext bit; for example, flipping a bit in the ciphertext causes the same bit to be flipped in the plaintext.

Self-synchronizing stream ciphers

Another approach uses several of the previous N ciphertext digits to compute the keystream. Such schemes are known as self-synchronizing stream ciphers, asynchronous stream ciphers or ciphertext autokey (CTAK). The idea of self-synchronization was patented in 1946 and has the advantage that the receiver will automatically synchronise with the keystream generator after receiving N ciphertext digits, making it easier to recover if digits are dropped or added to the message stream. Single-digit errors are limited in their effect, affecting only up to N plaintext digits.

An example of a self-synchronising stream cipher is a block cipher in cipher feedback (CFB) mode.

Based on linear-feedback shift registers

Binary stream ciphers are often constructed using linear-feedback shift registers (LFSRs) because they can be easily implemented in hardware and can be readily analysed mathematically. The use of LFSRs on their own, however, is insufficient to provide good security. Various schemes have been proposed to increase the security of LFSRs.

Non-linear combining functions

Because LFSRs are inherently linear, one technique for removing the linearity is to feed the outputs of several parallel LFSRs into a non-linear Boolean function to form a combination generator. Various properties of such a combining function are critical for ensuring the security of the resultant scheme, for example, in order to avoid correlation attacks.

Clock-controlled generators

Normally LFSRs are stepped regularly. One approach to introducing non-linearity is to have the LFSR clocked irregularly, controlled by the output of a second LFSR. Such generators include the stop-and-go generator, the alternating step generator and the shrinking generator.

An alternating step generator comprises three LFSRs, which we will call LFSR0, LFSR1 and LFSR2 for convenience. The output of one of the registers decides which of the other two is to be used; for instance, if LFSR2 outputs a 0, LFSR0 is clocked, and if it outputs a 1, LFSR1 is clocked instead. The output is the exclusive OR of the last bit produced by LFSR0 and LFSR1. The initial state of the three LFSRs is the key.

The stop-and-go generator (Beth and Piper, 1984) consists of two LFSRs. One LFSR is clocked if the output of a second is a 1, otherwise it repeats its previous output. This output is then (in some versions) combined with the output of a third LFSR clocked at a regular rate.

The shrinking generator takes a different approach. Two LFSRs are used, both clocked regularly. If the output of the first LFSR is 1, the output of the second LFSR becomes the output of the generator. If the first LFSR outputs 0, however, the output of the second is discarded, and no bit is output by the generator. This mechanism suffers from timing attacks on the second generator, since the first/control LFSR determines whether the second LFSR's bit is emitted or discarded. Thus the variable output rate leaks information about the control/selector sequence. This can be alleviated by buffering the output.

Filter generator

Another approach to improving the security of an LFSR is to pass the entire state of a single LFSR into a non-linear filtering function.

Other designs

Instead of a linear driving device, one may use a nonlinear update function. For example, Klimov and Shamir proposed triangular functions (T-functions) with a single cycle on n-bit words.

Security

For a stream cipher to be secure, its keystream must have a large period, and it must be impossible to recover the cipher's key or internal state from the keystream. Cryptographers also demand that the keystream be free of even subtle biases that would let attackers distinguish a stream from random noise, and free of detectable relationships between keystreams that correspond to related keys or related cryptographic nonces. That should be true for all keys (there should be no weak keys), even if the attacker can know or choose some plaintext or ciphertext.

As with other attacks in cryptography, stream cipher attacks can be certificational so they are not necessarily practical ways to break the cipher but indicate that the cipher might have other weaknesses.

Securely using a secure synchronous stream cipher requires that one never reuse the same keystream twice. That generally means a different nonce or key must be supplied to each invocation of the cipher. Application designers must also recognize that most stream ciphers provide not authenticity but privacy: encrypted messages may still have been modified in transit.

Short periods for stream ciphers have been a practical concern. For example, 64-bit block ciphers like DES can be used to generate a keystream in output feedback (OFB) mode. However, when not using full feedback, the resulting stream has a period of around 2³² blocks on average; for many applications, the period is far too low. For example, if encryption is being performed at a rate of 8 megabytes per second, a stream of period 2³² blocks will repeat after about an hour.

Some applications using the stream cipher RC4 are attackable because of weaknesses in RC4's key setup routine; new applications should either avoid RC4 or make sure all keys are unique and ideally unrelated (such as generated by a well-seeded CSPRNG or a cryptographic hash function) and that the first bytes of the keystream are discarded.

The elements of stream ciphers are often much simpler to understand than block ciphers and are thus less likely to hide any accidental or malicious weaknesses.

Usage

Stream ciphers are often used for their speed and simplicity of implementation in hardware, and in applications where plaintext comes in quantities of unknowable length like a secure wireless connection. If a block cipher (not operating in a stream cipher mode) were to be used in this type of application, the designer would need to choose either transmission efficiency or implementation complexity, since block ciphers cannot directly work on blocks shorter than their block size. For example, if a 128-bit block cipher received separate 32-bit bursts of plaintext, three quarters of the data transmitted would be padding. Block ciphers must be used in ciphertext stealing or residual block termination mode to avoid padding, while stream ciphers eliminate this issue by naturally operating on the smallest unit that can be transmitted (usually bytes).

Another advantage of stream ciphers in military cryptography is that the cipher stream can be generated in a separate box that is subject to strict security measures and fed to other devices such as a radio set, which will perform the XOR operation as part of their function. The latter device can then be designed and used in less stringent environments.

ChaCha is becoming the most widely used stream cipher in software; others include: RC4, A5/1, A5/2, Chameleon, FISH, Helix, ISAAC, MUGI, Panama, Phelix, Pike, Salsa20, SEAL, SOBER, SOBER-128, and WAKE.

Several stream ciphers have national origins or are mandated by specific national or international standards. Notable examples include ZUC (China), SNOW / SNOW 3G (Sweden, used in 3GPP standards alongside ZUC for 5G encryption), and the A5/1 and A5/2 algorithms (designed by ETSI for GSM). The eSTREAM portfolio finalists also have specific national origins, including Grain (Sweden), Trivium (Belgium), and Rabbit (Denmark).

Comparison