SKI is a novel cryptographic library that has been developed with a focus on hardware-constrained environments such as embedded systems, which are becoming more prevalent with the proliferation of "smart" devices and the Internet of Things. To date, security in such platforms has always been at odds with performance and design constraints, and thus suffered in comparison to its other features; such issues have precipitated widespread security breaches of unprecedented scale, such as the Mirai botnet (ca. 2016). When hardware and budgets are tight, economic incentives for security simply don't make the cut, especially in high-volume production runs. The SKI project thus aims to create a simple, interoperable, unified specification, using modern algorithms selected for speed (without sacrificing security), to ease the burden on embedded systems developers to introduce good security practices into their products from the start.
SKI is not the first cryptographic library. It borrows heavily from concepts introduced by Pretty Good Privacy (PGP, Zimmermann 1991), and is very similar in interface to D.J. Bernstein's 2012 "Networking and Cryptography Library" (NaCL), which has been revised ("forked") by other major projects, including libsodium (Frank Denis, 2013). Our major technical contribution is a focus on hardware constraints, such as processing performance, working memory, or the presence of a reliable entropy source. Design choices made in concession to such constraints will be mentioned throughout the presentation.
SKI provides three major cryptographic operations:
Symmetric ("shared-key") encryption: A scheme to protect the confidentiality of data where all trusted parties hold the exact same key;
Asymmetric ("public-key") key derivation: A scheme to derive a shared, trusted secret using two-part keys (a "keypair"): a public part that can be disclosed, and a private part kept secret;
Digital signature: A scheme to provide authenticity of data using a keypair, where the private part is present for attestation, and the public part can be used for verification.
For simplicity, SKI can use the same keypair for both key derivation and digital signature, but library users can separate this functionality amongst multiple keys if they so choose. SKI imposes no inherent limitations on the use of keys beyond that they are valid for the operation; such details, if they must be implemented, are delegated to the developer.
It is worth note that SKI does not provide "authenticated encryption" directly, while in most similar libraries, this is the only option. It is easy and, we believe, secure for the authentication and encryption primitives to be composed in any order to provide the same service. As an added benefit, the primitives are separable if the use case has no need for, or cannot perform well with, a particular one.
Perhaps the most innovative part of the design of this standard is the use of Uniform Resource Name (URN) syntax as specified in RFC 2141 (but note that we have not yet registered our namespace), along with RFC 3548 "URI-safe" Base64 encoding to ensure the binary data can be efficiently represented as text. This decision was primarily because this form is very easy to convey graphically--for example, in the ubiquitous "Quick Response" (QR) code. We anticipate that devices can come with a card or attachment (for example, a sticker), with such a graphical code conveying a SKI URN, and that this is a practical, scaleable, and secure way to do key exchange on a per-unit basis. (Such key exchanges may be used to bootstrap trust in the device and allow for further key exchange, or be the sole key on the device--we will discuss this more in the section on asymmetric keys and signatures).
All codable objects that SKI can use are compatible with this encoding, including variable-length data, such as encrypted data packets. However, the Base64 encoding, while compressible, is unwieldy in many circumstances, so SKI also allows for a "binary" encoding when using "8-bit clean" (pure binary) communication channels that will not be rendered as text. This is also a valid encoding for all objects, but it is generally used only when the packet size is linear in the input size (thus, encryption operations).
SKI data packets, in either URN or binary form, are not self-describing and require transport metadata, such as a packet length over a stream protocol (or a dedicated packet protocol) to communicate. We anticipate this is a reasonable burden to offload to library users, but if this proves to be false, the current coding standard is flexible enough to admit a self-describing packet.
Arguably the easiest form of encryption, symmetric encryption depends on a shared secret, referred to as a "symmetric key", or simply a key. A symmetric cipher is two operations, encryption and decryption, which take some amount of data and such a key, and for which encryption composed with decryption only with the same key is an identity. The design of such ciphers is considered secure when the key cannot efficiently be derived from the initial data (the plaintext), the encrypted data (the ciphertext), or both.
SKI uses the "XChaCha" variant of the ChaCha20 stream cipher, the latter by Bernstein (2008). As a stream cipher, the ciphertext and the plaintext have the same length, but--to prevent key recovery--each encryption with the same key must use a nonce (a number "used once"), which is safe to disclose but must be unique per key use. It is catastrophic to the security of the cryptosystem to use the same nonce with different plaintexts. The XChaCha variant we have chosen has a 256-bit key and 192-bit nonce, which should allow, in theory, an average of about 7.9e28 uniformly-randomly-chosen nonces to be used with the same key before such a failure occurs. However, it is unsafe to assume that embedded systems have a good source of entropy; we posit that it is safe to use a simple incrementing counter in non-volatile memory for the nonce on such constrained implementations, which allow for about 6.2e57 nonces before wrap-around occurs. It is an astronomically remote possibility that any single device will send this much data, even with a fixed key, over the course of its usable lifetime.
Our reference implementation provides the "ski:symk" scheme for the 256-bit key, and the "ski:syed" scheme for symmetrically-encrypted data, which is the simple concatenation of the 192-bit nonce and the ciphertext. Under the "sym" command:
A new, random key on a host with sufficient entropy can be generated with "gen".
A key derived from some input (such as a password) can be derived using "derive". We use the Argon2 key derivation function, which is specifically chosen to be hard against adversaries with access to FPGAs and difficult to implement on an ASIC with better performance than a modern computer. As it requires 8MB of random-access memory, we do not anticipate that such a command will be used on an embedded device; rather, only the key itself needs to be stored.
Ciphertext can be produced from plaintext using the "encrypt" command, and the inverse done with the "decrypt" command, both of which expect a symmetric key. Both use the binary encoding of the "syed" scheme by default, but can be given an argument to generate a URN instead.
Asymmetric encryption is so named for having two distinct keys for its operation, one of which can be disclosed. In a secure asymmetric cryptosystem, the disclosed "public" key cannot be used to efficiently derive the undisclosed "private" key in the same pair. Otherwise, the guarantees from a symmetric cipher hold analogously.
SKI uses Elliptic Curve Diffie-Hellman (ECDH) over Curve25519, an elliptic curve over the prime 2 ^ 255 - 19, developed by Bernstein (2005), which he conjectures to have the equivalent of 128-bit security (requires the equivalent of 2^128 brute-force attempts to break some aspect of the cryptosystem). The Diffie-Hellman protocol requires that both parties attempting to encrypt a message have disclosed their public keys to each other, and generates a shared secret key that is believed to be difficult to derive without access to at least one of the private keys. To preserve forward secrecy, the shared key, which is constant for the lifetime of the two keypairs, is used to encrypt a random per-message key; in this way, disclosure of the message key does not provide significant information about the shared key, or either private key. The cipher used for encrypting both the message key and the message itself is the selfsame symmetric cipher discussed previously.
In embedded implementations where randomness may be a concern, we posit that it is acceptable to use a non-volatile counter, as with the nonce, to generate the message keys. However, to avoid catastrophic loss of confidentiality, the initial values of the message key and symmetric nonce counters should be unrelated and never disclosed. We anticipate that it is acceptable to use some entropy at manufacture or program time to initialize the values of these counters independently, but for this reason, we encourage systems using asymmetric encryption to use a cryptographically-secure entropy source for these operations whenever resources allow (such as applications designed for general-purpose computers).
Our reference implementation provides the "ski:prvk" and "ski:pubk" URN schemes for private and public keys, respectively, both of which are 256-bit. It also provides the "ski:shak" scheme for the ECDH-derived shared key. Finally, it provides the "ski:encr" scheme for asymmetrically-encrypted data packets, which consists of the data of two "syed" packets: the first, of fixed length, is an encrypted message key, and the second, of variable length, is the message, encrypted with the message key. Under the "key" command:
A new, random private key on a host with sufficient entropy can be generated with "gen". Compared to many other contemporary cryptosystems, the generation of a private key for Curve25519 is relatively fast, and requires only 253 bits of entropy.
The public key for a given private key can be derived with "pub".
The shared key for one public and one private key can be generated with "shared". We don't recommend this in practice, because disclosing this key can result in loss of confidentiality in all messages between these two keys; it is provided as a diagnostic aid, and for integration with other systems.
Ciphertext can be produced from plaintext with "encrypt", and the inverse done with "decrypt". Both operations require a public key (usually the "recipient") and a private key (usually the "sender"). "decrypt" can recover using either the original private and public key, or (more typically) the private key of the recipient and the public key of the sender.
Future work includes extending the number of public keys to which a single message can be addressed, as it requires only constant size (the size of one encrypted message key) to add further recipients by public key.
A signature is a token, produced over some data, using a secret, which can be "verified" by other parties to ensure integrity of the data. For such a system to be secure, an adversary cannot efficiently derive such a token without the secret, nor derive the secret from the token, and any attempt to interfere with the integrity of the message (change its data in any way) or of the token should cause verification to fail.
Efficient cryptographic-hash-based schemes, known as Hash-based Message Authetication Codes (HMACs) exist for shared secrets, but are unimplemented in SKI. Instead, SKI uses the Elliptic Curve Digital Signature Algorithm, also over Bernstein's Curve25519, and using the 512-bit variant of the NIST Secure Hashing Algorithm 2, published 2001, as the message digest. Because Curve25519's field is 256-bit, we use only the first 256 bits of the 512 bit digest; nonetheless, SHA-512 has better preimage resistance than SHA-256 (the 256-bit variant), so we posit this design has better--or, at least, no worse--security properties.
As with other primitives, ECDSA requires a 255-bit "nonce" per signature; unlike the previous primitives, this "nonce" is a secret. Because of this change of role, we refer to this datum as a "nonce preimage". Disclosing the nonce preimage, or using the same nonce preimage for distinct messages, allows an adversary to recover the private key, thus defeating the integrity guarantees of this signature scheme and the confidentiality of any encrypted messages for which this private key was used. For this reason, we recommend using distinct signature and encryption keys whenever possible, to mitigate the damage caused by an accidental disclosure. While yet another nonce counter could be used for such nonces, SKI also provides for deterministic nonces, where the 255 bits are derived from a SHA-512 digest over the private key and the data. Thus, the security of this scheme relies on the resistance of the underlying hash function to collision; to the best of our knowledge, SHA-512 remains secure in this regard. Nonetheless, on platforms where access to entropy is readily available, we still recommend using randomized signatures, which relaxes the security proof to not rely on SHA-512's collision resistance.
Our reference implementation provides digital signatures through the "ski:sign" scheme, which consists of a "nonce postimage" (frequently labelled "r" in ECDSA literature) and the signature token (usually "s") derived from the private key and the given nonce preimage. Both of these are a constant 256 bits in length. Since the private and public keys are compatible with the asymmetric encryption keys, they can be used with two more "key" subcommands:
"sign", which generates a SKI signature for a message, and
"verify", which validates a SKI signature for a message.
Note that the SKI signature is of constant size and "detached"; it must be sent independently of the message.
Although not included as a primitive, we revisit that the two previous primitives combined--asymmetric encryption and digital signatures--can create "authenticated encryption". There are three ways, in general, to combine these functions:
Encrypt-then-MAC: Encrypt the data, then sign the ciphertext;
MAC-then-encrypt: Compute a signature, concatenate it with the plaintext, and encrypt the concatenation;
Encrypt-and-MAC: Sign the plaintext independently of encrypting it.
Various subtleties should be observed with each approach, but we believe that the systems we have chosen have beneficial properties regardless of AE scheme: primarily, the reliance on independent nonces makes it difficult to attack any scheme even if the same private key is used for both primitives, and the use of a stream cipher mitigates the effects of padding-oracle attacks. Note, however, that other protocol-specific oracles may be possible due to the malleability of the stream cipher, and thus we encourage authenticating encrypted data (using any above scheme) unless a compelling reason justifies otherwise.
Pragmatics and Miscellany
One familiar with other cryptosystems occupying a similar niche in software design, particularly PGP, will note that SKI has no concept of a "certificate". This is a simplifying assumption--we make no attempt to implement any kind of database, key storage, or key validation scheme, as we anticipate that these will depend heavily on implementation details and the environment in which a solution is deployed. Similarly, without certificates, SKI has no inbuilt concept of "key expiry"; it is up to the designers of devices to determine their rekeying policies, and requires support in the sense of a (trusted) timekeeping device, such as a real-time clock or networked time server. These, we believe, are ancillary to the design of a good foundation of a cryptosystem, though we have made every effort to ensure that the primitives provided by this library remain secure enough for practical use, as of the present, for the forseeable lifetimes of these embedded devices.
There is presently no quantum-resistant cryptography in SKI as of yet; the implementations of such cryptosystems we've reviewed so far have not met our standards of performance on embedded devices. However, this is only a pragmatic compromise for the time being; as hardware and cryptography improve, we plan on devoting future work to including quantum-resistant suites.
In summation, we present SKI, a simple, novel cryptographic library targeting embedded and hardware-constrained systems, with a focus on performance, portability, and ease of use. We discussed its cryptographic primitives, their usage, and the concessions made pursuant these goals. We release an implementation freely, in hopes that it can improve the status quo of security for embedded platforms and devices as they continue to proliferate, and we hope that this contribution will help abate the security concerns of such technologies.
Thank you for listening.