diff --git a/draft-ietf-pquip-pqc-engineers.md b/draft-ietf-pquip-pqc-engineers.md index e735384..1d8cc0c 100644 --- a/draft-ietf-pquip-pqc-engineers.md +++ b/draft-ietf-pquip-pqc-engineers.md @@ -278,24 +278,24 @@ The candidates still advancing for standardization are: * [Classic McEliece](https://classic.mceliece.org/): Based on the hardness of syndrome decoding of Goppa codes. Goppa codes are a class of error-correcting codes that can correct a certain number of errors in a transmitted message. The decoding problem involves recovering the original message from the received noisy codeword. * [BIKE](https://bikesuite.org/): Based on the the hardness of syndrome decoding of QC-MDPC codes. Quasi-Cyclic Moderate Density Parity Check (QC-MDPC) code are a class of error correcting codes that leverages bit flipping technique to efficiently correct errors. -* [HQC](http://pqc-hqc.org/) : Based on the hardness of syndrome decoding of Quasi-cyclic concatenated Reed Muller Reed Solomon (RMRS) codes in the Hamming metric. Reed Muller (RM) codes are a class of block error correcting codes used especially in wireless and deep space communications. Reed Solomon (RS) are a class of block error correcting codes that are used to detect and correct multiple bit errors. +* [HQC](http://pqc-hqc.org/): Based on the hardness of syndrome decoding of Quasi-cyclic concatenated Reed Muller Reed Solomon (RMRS) codes in the Hamming metric. Reed Muller (RM) codes are a class of block error correcting codes used especially in wireless and deep space communications. Reed Solomon (RS) are a class of block error correcting codes that are used to detect and correct multiple bit errors. * [SIKE](https://sike.org/) (Broken): Supersingular Isogeny Key Encapsulation (SIKE) is a specific realization of the SIDH (Supersingular Isogeny Diffie-Hellman) protocol. Recently, a [mathematical attack](https://eprint.iacr.org/2022/975.pdf) based on the "glue-and-split" theorem from 1997 from Ernst Kani was found against the underlying chosen starting curve and torsion information. In practical terms, this attack allows for the efficient recovery of the private key. NIST announced that SIKE was no longer under consideration, but the authors of SIKE had asked for it to remain in the list so that people are aware that it is broken. While SIKE is broken, Isogenies in general remain an active area of cryptographic research due to their very attractive bandwidth usage, and we may yet see more cryptographic primitives in the future from this research area. # Threat of CRQCs on Cryptography Post-quantum cryptography or quantum-safe cryptography refers to cryptographic algorithms that are secure against cryptographic attacks from both CRQCs and classic computers. -When considering the security risks associated with the ability of a quantum computer to attack traditional cryptography, it is important to distinguish between the impact on symmetric algorithms and public-key ones. Dr. Peter Shor and Dr. Lov Grover developed two algorithms that changed the way the world thinks of security under the presence of a CRQC. +When considering the security risks associated with the ability of a quantum computer to attack traditional cryptography, it is important to distinguish between the impact on symmetric algorithms and public-key ones. Dr. Peter Shor and Dr. Lov Grover developed two algorithms that changed the way the world thinks of security in the presence of a CRQC. -It is also worth some discussion of the term "quantum adversary". Quantum computers are, by their nature, hybrids of classical and quantum computational units. For example, Shor's algorithm consists of a combination of quantum and classical computational steps. Thus, the term "quantum adversary" should be thought of as 'quantum-enhanced adversary,' meaning they have access to both classical and quantum computational techniques. +It is also worth discussing the term "quantum adversary". Quantum computers are, by their nature, hybrids of classical and quantum computational units. For example, Shor's algorithm consists of a combination of quantum and classical computational steps. Thus, the term "quantum adversary" should be thought of as "quantum-enhanced adversary," meaning they have access to both classical and quantum computational techniques. -Despite the fact that large-scale quantum computers do not yet exist to experiment on, the theoretical properties of quantum computation are very well understood. This allows us to reason today about the upper limits of quantum-enhanced computation, and indeed to design cryptographic algorithms that are resistant to any conceivable for of quantum cryptanalysis. +Despite the fact that large-scale quantum computers do not yet exist to experiment on, the theoretical properties of quantum computation are very well understood. This allows us to reason today about the upper limits of quantum-enhanced computation, and indeed to design cryptographic algorithms that are resistant to any conceivable form of quantum cryptanalysis. ## Symmetric cryptography {#symmetric} For unstructured data such as symmetric encrypted data or cryptographic hashes, although CRQCs can search for specific solutions across all possible input combinations (e.g., Grover's Algorithm), no quantum algorithm is known to break the underlying security properties of these classes of algorithms. -Grover's algorithm is a quantum search algorithm that provides a theoretical quadratic speedup for searching an unstructured database, compared to classical search algorithms. If we consider the mapping of hash values to their corresponding hash inputs (also known as pre-image), or of ciphertext blocks to the corresponding plaintext blocks, as an unstructured database, then Grover’s algorithm theoretically requires doubling the key sizes of the symmetric algorithms that are currently deployed today to counter the quadratic speedup and maintain current security level. This is because Grover’s algorithm reduces the amount of operations to break 128-bit symmetric cryptography to 2^{64} quantum operations, which might sound computationally feasible. However, 2^{64} operations performed in parallel are feasible for modern classical computers, but 2^{64} quantum operations performed serially in a quantum computer are not. Grover's algorithm is highly non-parallelizable and even if one deploys 2^c computational units in parallel to brute-force a key using Grover's algorithm, it will complete in time proportional to 2^{(128−c)/2}, or, put simply, using 256 quantum computers will only reduce runtime by a factor of 16, 1024 quantum computers will only reduce runtime by a factor of 32 and so forth ​(see {{NIST}} and {{Cloudflare}}​). Therefore, while Grover's attack suggests that we should double the sizes of symmetric keys, the current consensus among experts is that the current key sizes remain secure in practice. +Grover's algorithm is a quantum search algorithm that provides a theoretical quadratic speedup for searching an unstructured database, compared to classical search algorithms. If we consider the mapping of hash values to their corresponding hash inputs (also known as pre-image), or of ciphertext blocks to the corresponding plaintext blocks, as an unstructured database, then Grover’s algorithm theoretically requires doubling the key sizes of the symmetric algorithms that are currently deployed today to counter the quadratic speedup and maintain current security levels. This is because Grover’s algorithm reduces the amount of operations to break 128-bit symmetric cryptography to 2^{64} quantum operations, which might sound computationally feasible. However, 2^{64} operations performed in parallel are feasible for modern classical computers, but 2^{64} quantum operations performed serially in a quantum computer are not. Grover's algorithm is highly non-parallelizable and even if one deploys 2^c computational units in parallel to brute-force a key using Grover's algorithm, it will complete in time proportional to 2^{(128−c)/2}, or, put simply, using 256 quantum computers will only reduce runtime by a factor of 16, 1024 quantum computers will only reduce runtime by a factor of 32 and so forth ​(see {{NIST}} and {{Cloudflare}}​). Therefore, while Grover's attack suggests that we should double the sizes of symmetric keys, the current consensus among experts is that the current key sizes remain secure in practice. How can someone be sure that an improved algorithm won’t outperform Grover's algorithm at some point in time? Christof Zalka has shown that Grover's algorithm (and in particular its non-parallel nature) achieves the best possible complexity for unstructured search {{Grover-search}}. @@ -303,7 +303,7 @@ Finally, in their evaluation criteria for PQC, NIST is assessing the security le ## Asymmetric cryptography -“Shor’s algorithm” on the other side, efficiently solves the integer factorization problem (and the related discrete logarithm problem), which offer the foundations of the vast majority of public-key cryptography that the world uses today. This implies that, if a CRQC is developed, today’s public-key cryptography algorithms (e.g., RSA, Diffie-Hellman and Elliptic Curve Cryptography, as well as less commonly-used variants such as ElGamal and Schnorr signatures) and protocols would need to be replaced by algorithms and protocols that can offer cryptanalytic resistance against CRQCs. Note that Shor’s algorithm cannot run solely on a classic computer, it needs a CRQC. +“Shor’s algorithm” on the other hand, efficiently solves the integer factorization problem (and the related discrete logarithm problem), which offer the foundations of the vast majority of public-key cryptography that the world uses today. This implies that, if a CRQC is developed, today’s public-key cryptography algorithms (e.g., RSA, Diffie-Hellman and Elliptic Curve Cryptography, as well as less commonly-used variants such as ElGamal and Schnorr signatures) and protocols would need to be replaced by algorithms and protocols that can offer cryptanalytic resistance against CRQCs. Note that Shor’s algorithm cannot run solely on a classic computer, it needs a CRQC. For example, to provide some context, one would need 20 million noisy qubits to break RSA-2048 in 8 hours {{RSAShor}}{{RSA8HRS}} or 4099 stable (or logical) qubits to break it {{RSA10SC}}. @@ -317,9 +317,9 @@ The field of cryptographic side-channel attacks potentially stands to gain a boo The timeline, and driving motivation for transition differs slightly between data confidentiality (e.g., encryption) and data authentication (e.g., signature) use-cases. -For data confidentiality, we are concerned with the so-called "Harvest Now, Decrypt Later" attack where a malicious actor with adequate resources can launch an attack to store sensitive encrypted data today that can be decrypted once a CRQC is available. This implies that, every day, sensitive encrypted data is susceptible to the attack by not implementing quantum-safe strategies, as it corresponds to data being deciphered in the future. +For data confidentiality, we are concerned with the so-called "Harvest Now, Decrypt Later" attack where a malicious actor with adequate resources can launch an attack to store sensitive encrypted data today that they hope to decrypt once a CRQC is available. This implies that, every day, sensitive encrypted data is susceptible to the attack by not implementing quantum-safe strategies, as it corresponds to data possibly being deciphered in the future. -For authentication, it is often the case that signatures have a very short lifetime between signing and verifying -- such as during a TLS handshake -- but some authentication use-cases do require long lifetimes, such as signing firmware or software that will be active for decades, signing legal documents, or signing certificates that will be embedded into hardware devices such as smartcards. And even for short-lived signatures use cases, the infrastructure often relies on long-lived root keys which can be difficult to update or replace on in-field devices. +For authentication, it is often the case that signatures have a very short lifetime between signing and verifying -- such as during a TLS handshake -- but some authentication use-cases do require long lifetimes, such as signing firmware or software that will be active for decades, signing legal documents, or signing certificates that will be embedded into hardware devices such as smartcards. Even for short-lived signatures use cases, the infrastructure often relies on long-lived root keys which can be difficult to update or replace on in-field devices. ~~~~~ @@ -334,11 +334,11 @@ For authentication, it is often the case that signatures have a very short lifet ~~~~~ {: #Mosca title="Mosca model"} -These challenges are illustrated nicely by the so-called Mosca model discussed in ​{{Threat-Report}}. In the {{Mosca}}, "x" denotes the time that our systems and data need to remain secure, "y" the number of years to fully migrate to a PQC infrastructure and "z" the time until a CRQC that can break current cryptography is available. The model assumes either that encrypted data can be intercepted and stored before the migration is completed in "y" years, or that signatures will still be relied upon for "x" years after their creation. This data remains vulnerable for the complete "x" years of their lifetime, thus the sum "x+y" gives us an estimate of the full timeframe that data remain insecure. The model essentially asks how are we preparing our IT systems during those "y" years (or in other words, how can one minimize those "y" years) to minimize the transition phase to a PQC infrastructure and hence minimize the risks of data being exposed in the future. +These challenges are illustrated nicely by the so-called Mosca model discussed in ​{{Threat-Report}}. In the {{Mosca}}, "x" denotes the time that our systems and data need to remain secure, "y" the number of years to fully migrate to a PQC infrastructure and "z" the time until a CRQC that can break current cryptography is available. The model assumes either that encrypted data can be intercepted and stored before the migration is completed in "y" years, or that signatures will still be relied upon for "x" years after their creation. This data remains vulnerable for the complete "x" years of their lifetime, thus the sum "x+y" gives us an estimate of the full timeframe that data remain insecure. The model essentially asks how we are preparing our IT systems during those "y" years (in other words, how one can minimize those "y" years) to minimize the transition phase to a PQC infrastructure and hence minimize the risks of data being exposed in the future. Finally, other factors that could accelerate the introduction of a CRQC should not be under-estimated, like for example faster-than-expected advances in quantum computing and more efficient versions of Shor’s algorithm requiring fewer qubits. Innovation often comes in waves, so it is to the industry’s benefit to remain vigilant and prepare as early as possible. Bear in mind also that while we track advances from public research institutions such as universities and companies that publish their results, there is also a great deal of large-budget quantum research being conducted privately by various national interests. Therefore, the true state of quantum computer advancement is likely several years ahead of the publicly available research. -Organizations should also consider carefully and honestly what their migration timeline "y" actually is. If you think only of the time between receiving a patch from your technology vendor, and rolling that patch out, then "y" might seem as short as a few weeks. However, this represents the minority of migration cases; more often a PQC migration will involve at least some amount of hardware replacement. For example performance-sensitive applications will need CPUs with PQC hardware acceleration. Security-sensitive applications will need PQC TPMs, TEEs, Secure Enclaves, and other cryptographic co-processors. Smartcard applications will require replacement of the cards and also of the readers which can come in many form-factors: tap-for-entry door and turnstile readers, PIN pad machines, laptops with built-in smartcard readers, and many others. Included in "y" is not only the deployment time, but also preparation time: integration, testing, auditing and re-certification of cryptographic environments. Consider also upstream effects that contribute to "y", including lead-times for your vendors to produce PQC-ready products, which may itself include auditing and certification delays, time for regulating bodies to adopt PQC policies, time for auditors to become familiar with the new requirements, etc. If you measure the full migration time "y" from when your vendors begin implementing PQC functionality, to when you switch off your last non-PQC-capable device, then "y" can be quite long; likely measured in years or decades for most moderately-sized organizations. +Organizations should also consider carefully and honestly what their migration timeline "y" actually is. If you think only of the time between receiving a patch from your technology vendor, and rolling that patch out, then "y" might seem as short as a few weeks. However, this represents the minority of migration cases; more often, a PQC migration will involve at least some amount of hardware replacement. For example, performance-sensitive applications will need CPUs with PQC hardware acceleration. Security-sensitive applications will need PQC TPMs, TEEs, Secure Enclaves, and other cryptographic co-processors. Smartcard applications will require replacement of the cards as well as of the readers which can come in many form-factors: tap-for-entry door and turnstile readers, PIN pad machines, laptops with built-in smartcard readers, and many others. Included in "y" is not only the deployment time, but also preparation time: integration, testing, auditing, and re-certification of cryptographic environments. Consider also upstream effects that contribute to "y", including lead-times for your vendors to produce PQC-ready products, which may itself include auditing and certification delays, time for regulating bodies to adopt PQC policies, time for auditors to become familiar with the new requirements, etc. If you measure the full migration time "y" from when your vendors begin implementing PQC functionality, to when you switch off your last non-PQC-capable device, then "y" can be quite long; likely measured in years or decades for even most moderately-sized organizations. # Post-quantum cryptography categories @@ -348,19 +348,19 @@ The current set of problems used in post-quantum cryptography can be currently g Lattice-based public-key cryptography leverages the simple construction of lattices (i.e., a regular collection of points in a Euclidean space that are evenly spaced) to create 'trapdoor' problems. These problems are efficient to compute if you possess the secret information but challenging to compute otherwise. Examples of such problems include the Shortest Vector, Closest Vector, Shortest Integer Solution, Learning with Errors, Module Learning with Errors, and Learning with Rounding problems. All of these problems feature strong proofs for worst-to-average case reduction, effectively relating the hardness of the average case to the worst case. -The possibility to implement public-key schemes on lattices is tied to the characteristics of the vector basis used for the lattice. In particular, solving any of the mentioned problems can be easy when using "reduced" or "good" bases (i.e., as short as possible and as orthogonal as possible), while it becomes computationally infeasible when using "bad" bases (i.e., long vectors not orthogonal). Although the problem might seem trivial, it is computationally hard when considering many dimensions, or when the underlying field is not simple numbers, but high-order polynomials. Therefore, a typical approach is to use "bad" basis for public keys and "good" basis for private keys. The public keys ("bad" basis) let you easily verify signatures by checking, for example, that a vector is the closest or smallest, but do not let you solve the problem (i.e., finding the vector) that would yield the private key. Conversely, private keys (i.e., the "good" basis) can be used for generating the signatures (e.g., finding the specific vector). +The possibility to implement public-key schemes on lattices is tied to the characteristics of the vector basis used for the lattice. In particular, solving any of the mentioned problems can be easy when using "reduced" or "good" bases (i.e., as short as possible and as orthogonal as possible), while it becomes computationally infeasible when using "bad" bases (i.e., long, non-orthogonal vectors). Although the problem might seem trivial, it is computationally hard when considering many dimensions, or when the underlying field is not simple numbers, but high-order polynomials. Therefore, a typical approach is to use "bad" bases for public keys and "good" bases for private keys. The public keys ("bad" bases) let you easily verify signatures by checking, for example, that a vector is the closest or smallest, but do not let you solve the problem (i.e., finding the vector) that would yield the private key. Conversely, private keys (i.e., the "good" bases) can be used for generating the signatures (e.g., finding the specific vector). -Lattice-based schemes usually have good performances and average size public keys and signatures (average within the PQC primitives at least, they are still several orders of magnitude larger than RSA or ECC signatures), making them the best available candidates for general-purpose use such as replacing the use of RSA in PKIX certificates. +Lattice-based schemes usually have good performances and average size public keys and signatures (average within the PQC primitives at least; they are still several orders of magnitude larger than e.g., RSA or ECC signatures), making them the best available candidates for general-purpose use such as replacing the use of RSA in PKIX certificates. -Examples of such class of algorithms include ML-KEM, FN-DSA and ML-DSA. +Examples of this class of algorithms include ML-KEM, FN-DSA and ML-DSA. It is noteworthy that lattice-based encryption schemes require a rounding step during decryption which has a non-zero probability of "rounding the wrong way" and leading to a decryption failure, meaning that valid encryptions are decrypted incorrectly; as such, an attacker could significantly reduce the security of lattice-based schemes that have a relatively high failure rate. However, for most of the NIST Post-Quantum Proposals, the number of required oracle queries to force a decryption failure is above practical limits, as has been shown in {{LattFail1}}. More recent works have improved upon the results in {{LattFail1}}, showing that the cost of searching for additional failing ciphertexts after one or more have already been found, can be sped up dramatically {{LattFail2}}. Nevertheless, at this point in time (July 2023), the PQC candidates by NIST are considered secure under these attacks and we suggest constant monitoring as cryptanalysis research is ongoing. ## Hash-Based Public-Key Cryptography {#hash-based} -Hash based PKC has been around since the 1970s, when it was developed by Lamport and Merkle. It is used to create digital signature algorithms and its security is mathematically based on the security of the selected cryptographic hash function. Many variants of hash-based signatures (HBS) have been developed since the 70s including the recent XMSS {{!RFC8391}}, HSS/LMS {{!RFC8554}} or BPQS schemes. Unlike digital signature techniques, most hash-based signature schemes are stateful, which means that signing necessitates the update and careful tracking of the secret key. Producing multiple signatures using the same secret key state results in loss of security and ultimately signature forgery attacks against that key. +Hash based PKC has been around since the 1970s, when it was developed by Lamport and Merkle. It is used to create digital signature algorithms and its security is mathematically based on the security of the selected cryptographic hash function. Many variants of hash-based signatures (HBS) have been developed since the 70s including the recent XMSS {{!RFC8391}}, HSS/LMS {{!RFC8554}} or BPQS schemes. Unlike digital signature techniques, most hash-based signature schemes are stateful, which means that signing necessitates the update and careful tracking of the secret key. Producing multiple signatures using the same secret key state results in loss of security and may ultimately enable signature forgery attacks against that key. -Stateful hash-based signatures with long service lifetimes require additional operational complexity compared with other signature types. For example, consider a 20-year root key; there is an expectation that 20 years is longer than the expected lifetime of the hardware that key is stored on, and therefore the key will need to be migrated to new hardware at some point. Disaster-recovery scenarios where the primary node fail without warning can be similarly tricky. This requires careful operational and compliance consideration to ensure that no private key state can be re-used across the migration or disaster recovery event. One approach for avoiding these issues is to only use stateful HBS for short-term use cases that do not require horizontal scaling, for example signing a batch of firmware images and then retiring the signing key. +Stateful hash-based signatures with long service lifetimes require additional operational complexity compared with other signature types. For example, consider a 20-year root key; there is an expectation that 20 years is longer than the expected lifetime of the hardware that key is stored on, and therefore the key will need to be migrated to new hardware at some point. Disaster-recovery scenarios where the primary node fails without warning can be similarly tricky. This requires careful operational and compliance consideration to ensure that no private key state can be re-used across the migration or disaster recovery event. One approach for avoiding these issues is to only use stateful HBS for short-term use cases that do not require horizontal scaling, for example signing a batch of firmware images and then retiring the signing key. The SLH-DSA algorithm on the other hand leverages the HORST (Hash to Obtain Random Subset with Trees) technique and remains the only hash based signature scheme that is stateless, thus avoiding all the complexities with state management. @@ -376,9 +376,9 @@ Examples include all the NIST Round 4 (unbroken) finalists: Classic McEliece, HQ ## What is a KEM -A Key Encapsulation Mechanism (KEM) is a cryptographic technique used for securely exchanging symmetric key material between two parties over an insecure channel. It is commonly used in hybrid encryption schemes, where a combination of asymmetric (public key) and symmetric encryption is employed. The KEM encapsulation results in a fixed-length symmetric key that can be used with a symmetric algorithm, typically a block cipher, in one of two ways: (1) Derive a Data Encryption Key (DEK) to encrypt the data (2) Derive a Key Encryption Key (KEK) used to wrap a DEK. These techniques are often referred to as "hybrid public key encryption (HPKE)" {{?RFC9180}} mechanism. +A Key Encapsulation Mechanism (KEM) is a cryptographic technique used for securely exchanging symmetric key material between two parties over an insecure channel. It is commonly used in hybrid encryption schemes, where a combination of asymmetric (public key) and symmetric encryption is employed. The KEM encapsulation results in a fixed-length symmetric key that can be used with a symmetric algorithm, typically a block cipher, in one of two ways: (1) Derive a Data Encryption Key (DEK) to encrypt the data, or (2) Derive a Key Encryption Key (KEK) used to wrap a DEK. These techniques are often referred to as "hybrid public key encryption (HPKE)" {{?RFC9180}} mechanism. -The term "encapsulation" is chosen intentionally to indicate that KEM algorithms behave differently at the API level than the Key Agreement or Key Encipherment / Key Transport mechanisms that we are accustomed to using today. Key Agreement schemes imply that both parties contribute a public / private keypair to the exchange, while Key Encipherment / Key Transport schemes imply that the symmetric key material is chosen by one party and "encrypted" or "wrapped" for the other party. KEMs, on the other hand, behave according to the following API: +The term "encapsulation" is chosen intentionally to indicate that KEM algorithms behave differently at the API level from the Key Agreement or Key Encipherment / Key Transport mechanisms that we are accustomed to using today. Key Agreement schemes imply that both parties contribute a public / private keypair to the exchange, while Key Encipherment / Key Transport schemes imply that the symmetric key material is chosen by one party and "encrypted" or "wrapped" for the other party. KEMs, on the other hand, behave according to the following API: KEM relies on the following primitives [PQCAPI]: @@ -448,7 +448,7 @@ Authenticated Key Exchange with KEMs where both parties contribute a KEM public ~~~~~ {: #tab-dh-ake title="Diffie-Hellman based Authenticated Key Exchange"} -What's important to note about the sample flow above is that the shared secret `ss` is derived using key material from both the Client and the Server, which classifies it as an Authenticated Key Exchange (AKE). There is another property of a key exchange, called Non-Interactive Key Exchange (NIKE) which refers to whether the sender can compute the shared secret `ss` and encrypting content without requiring active interaction -- ie an exchange of network messages -- with the recipient. {{tab-dh-ake}} shows a Diffie-Hellman key exchange which is an AKE, since both parties are using long-term keys which can have established trust for example via certificates, but it is not a NIKE since the client needs to wait for the network interaction to receive the receiver's public key `pk2` before it can compute the shared secret `ss` and begin content encryption. However, a DH key exchange can be an AKE and a NIKE at the same time if the receiver's public key is known to the sender in advance, and many Internet Protocols rely on this property of DH-based key exchanges. +What's important to note about the sample flow above is that the shared secret `ss` is derived using key material from both the Client and the Server, which classifies it as an Authenticated Key Exchange (AKE). There is another property of a key exchange, called Non-Interactive Key Exchange (NIKE) which refers to whether the sender can compute the shared secret `ss` and encrypting content without requiring active interaction -- ie an exchange of network messages -- with the recipient. {{tab-dh-ake}} shows a Diffie-Hellman key exchange which is an AKE, since both parties are using long-term keys which can have established trust (for example, via certificates), but it is not a NIKE, since the client needs to wait for the network interaction to receive the receiver's public key `pk2` before it can compute the shared secret `ss` and begin content encryption. However, a DH key exchange can be an AKE and a NIKE at the same time if the receiver's public key is known to the sender in advance, and many Internet Protocols rely on this property of DH-based key exchanges. ~~~~~ aasvg +---------+ +---------+ @@ -476,7 +476,7 @@ What's important to note about the sample flow above is that the shared secret ` ~~~~~ {: #tab-dh-ake-nike title="Diffie-Hellman based Authenticated Key Exchange and Non-Interactive Key Exchange simultaneously"} -The complication with KEMs is that a KEM `Encaps()` is non-deterministic; it involves randomness chosen by the sender of that KEM. Therefore, in order to perform an AKE, the client must wait for the server to generate the needed randomness and perform `Encaps()` against the client key, which necessarily requires a network round-trip. Therefore a KEM-based protocol can either be an AKE or a NIKE, but cannot be both at the same time. Consequently, certain Internet protocols will necessitate redesign to accommodate this distinction, either by introducing extra network round-trips or by making trade-offs in security properties. +The complication with KEMs is that a KEM `Encaps()` is non-deterministic; it involves randomness chosen by the sender of that KEM. Therefore, in order to perform an AKE, the client must wait for the server to generate the needed randomness and perform `Encaps()` against the client key, which necessarily requires a network round-trip. Therefore, a KEM-based protocol can either be an AKE or a NIKE, but cannot be both at the same time. Consequently, certain Internet protocols will necessitate a redesign to accommodate this distinction, either by introducing extra network round-trips or by making trade-offs in security properties. ~~~~~ aasvg +---------+ +---------+ @@ -510,15 +510,15 @@ The complication with KEMs is that a KEM `Encaps()` is non-deterministic; it inv ~~~~~ {: #tab-kem-ake title="KEM based Authenticated Key Exchange"} -Here, `Combiner(ss1, ss2)`, often referred to as a KEM Combiner is a cryptographic construction that takes in two shared secrets and returns a single combined shared secret. The simplest combiner is concatenation `ss1 || ss2`, but combiners can vary in complexity depending on the cryptographic properties required. For example if the combination should preserve IND-CCA2 of either input even if the other is chosen maliciously, then a more complex construct is required. Another consideration for combiner design is so-called "binding properties" introduced in [KEEPINGUP] which may require the ciphertexts and recipient public keys to be included in the combiner. KEM combiner security analysis becomes more complicated in hybrid settings where the two KEMs represent different algorithms, for example one is ML-KEM and the other is ECDHE. For a more thorough discussion of KEM combiners, see [KEEPINGUP], {{?I-D.draft-ounsworth-cfrg-kem-combiners-04}}, and {{?I-D.draft-connolly-cfrg-xwing-kem-02}}. +Here, `Combiner(ss1, ss2)`, often referred to as a KEM Combiner, is a cryptographic construction that takes in two shared secrets and returns a single combined shared secret. The simplest combiner is concatenation `ss1 || ss2`, but combiners can vary in complexity depending on the cryptographic properties required. For example, if the combination should preserve IND-CCA2 of either input even if the other is chosen maliciously, then a more complex construct is required. Another consideration for combiner design is so-called "binding properties" introduced in [KEEPINGUP], which may require the ciphertexts and recipient public keys to be included in the combiner. KEM combiner security analysis becomes more complicated in hybrid settings where the two KEMs represent different algorithms, for example, where one is ML-KEM and the other is ECDHE. For a more thorough discussion of KEM combiners, see [KEEPINGUP], {{?I-D.draft-ounsworth-cfrg-kem-combiners-04}}, and {{?I-D.draft-connolly-cfrg-xwing-kem-02}}. ## Security properties ### IND-CCA2 -IND-CCA2 : IND-CCA2 (INDistinguishability under adaptive Chosen-Ciphertext Attack) is an advanced security notion for encryption schemes. It ensures the confidentiality of the plaintext and resistance against chosen-ciphertext attacks. An appropriate definition of IND-CCA2 security for KEMs can be found in [CS01] and [BHK09]. ML-KEM [ML-KEM] and Classic McEliece provide IND-CCA2 security. +IND-CCA2 (INDistinguishability under adaptive Chosen-Ciphertext Attack) is an advanced security notion for encryption schemes. It ensures the confidentiality of the plaintext and resistance against chosen-ciphertext attacks. An appropriate definition of IND-CCA2 security for KEMs can be found in [CS01] and [BHK09]. ML-KEM [ML-KEM] and Classic McEliece provide IND-CCA2 security. -Understanding IND-CCA2 security is essential for individuals involved in designing or implementing cryptographic systems and protocols in order to evaluate the strength of the algorithm, assess its suitability for specific use cases, and ensure that data confidentiality and security requirements are met. Understanding IND-CCA2 security is generally not necessary for developers migrating to using an IETF-vetted key establishment method (KEM) within a given protocol or flow. IND-CCA2 is considered the highest bar that a public key encryption mechanism can meet, and therefore is suitable for all uses. IETF specification authors should include all security concerns in the 'Security Considerations' section of the relevant RFC and not rely on implementers being deep experts in cryptographic theory. +Understanding IND-CCA2 security is essential for individuals involved in designing or implementing cryptographic systems and protocols in order to evaluate the strength of the algorithm, assess its suitability for specific use cases, and ensure that data confidentiality and security requirements are met. Understanding IND-CCA2 security is generally not necessary for developers migrating to using an IETF-vetted key establishment method (KEM) within a given protocol or flow. IND-CCA2 is considered the highest bar that a public key encryption mechanism can meet, and therefore is suitable for all uses. IETF specification authors should include all security concerns in the 'Security Considerations' section of the relevant RFC and not rely on implementers being experts in cryptographic theory. ### Binding @@ -536,7 +536,7 @@ HPKE (Hybrid Public Key Encryption) {{?RFC9180}} is a specific instantiation of ## What is a Post-quantum Signature -Any digital signature scheme that provides a construction defining security under post-quantum setting falls under this category of PQ signatures. +Any digital signature scheme that provides a construction defining security under a post-quantum setting falls under this category of PQ signatures. ## Security properties @@ -544,7 +544,7 @@ Any digital signature scheme that provides a construction defining security unde EUF-CMA (Existential Unforgeability under Chosen Message Attack) [GMR88] is a security notion for digital signature schemes. It guarantees that an adversary, even with access to a signing oracle, cannot forge a valid signature for an arbitrary message. EUF-CMA provides strong protection against forgery attacks, ensuring the integrity and authenticity of digital signatures by preventing unauthorized modifications or fraudulent signatures. ML-DSA, FN-DSA and SLH-DSA provide EUF-CMA security. -Understanding EUF-CMA security is essential for individuals involved in designing or implementing cryptographic systems in order to ensure the security, reliability, and trustworthiness of digital signature schemes. It allows for informed decision-making, vulnerability analysis, compliance with standards, and designing systems that provide strong protection against forgery attacks. Understanding EUF-CMA security is generally not necessary for developers migrating to using an IETF-vetted post-quantum cryptography (PQC) signature scheme within a given protocol or flow. EUF-CMA is considered the highest bar that a public key signature algorithm can meet, and therefore is suitable for all uses. IETF specification authors should include all security concerns in the 'Security Considerations' section of the relevant RFC and should not assume that implementers are deep experts in cryptographic theory +Understanding EUF-CMA security is essential for individuals involved in designing or implementing cryptographic systems in order to ensure the security, reliability, and trustworthiness of digital signature schemes. It allows for informed decision-making, vulnerability analysis, compliance with standards, and designing systems that provide strong protection against forgery attacks. Understanding EUF-CMA security is generally not necessary for developers migrating to using an IETF-vetted post-quantum cryptography (PQC) signature scheme within a given protocol or flow. EUF-CMA is considered the highest bar that a public key signature algorithm can meet, and therefore is suitable for all uses. IETF specification authors should include all security concerns in the 'Security Considerations' section of the relevant RFC and should not assume that implementers are experts in cryptographic theory. ## Details of FN-DSA, ML-DSA, and SLH-DSA {#sig-scheme} @@ -552,9 +552,9 @@ ML-DSA [ML-DSA] is a digital signature algorithm (part of the CRYSTALS suite) ba ML-DSA offers both deterministic and randomized signing and is instantiated with 3 parameter sets providing different security levels. Security properties of ML-DSA are discussed in Section 9 of {{?I-D.ietf-lamps-dilithium-certificates}}. -FN-DSA [FN-DSA] is based on the GPV hash-and-sign lattice-based signature framework introduced by Gentry, Peikert and Vaikuntanathan, which is a framework that requires a certain class of lattices and a trapdoor sampler technique. +FN-DSA [FN-DSA] is based on the GPV hash-and-sign lattice-based signature framework introduced by Gentry, Peikert, and Vaikuntanathan, which is a framework that requires a certain class of lattices and a trapdoor sampler technique. -The main design principle of FN-DSA is compactness, i.e. it was designed in a way that achieves minimal total memory bandwidth requirement (the sum of the signature size plus the public key size). This is possible due to the compactness of NTRU lattices. FN-DSA also offers very efficient signing and verification procedures. The main potential downsides of FN-DSA refer to the non-triviality of its algorithms and the need for floating point arithmetic support in order to support Gaussian-distributed random number sampling where the other lattice schemes use the less efficient but easier to support uniformly-distributed random number sampling. +The main design principle of FN-DSA is compactness, i.e., it was designed in a way that achieves minimal total memory bandwidth requirement (the sum of the signature size plus the public key size). This is possible due to the compactness of NTRU lattices. FN-DSA also offers very efficient signing and verification procedures. The main potential downsides of FN-DSA refer to the non-triviality of its algorithms and the need for floating point arithmetic support in order to support Gaussian-distributed random number sampling where the other lattice schemes use the less efficient but easier to support uniformly-distributed random number sampling. Implementers of FN-DSA need to be aware that FN-DSA signing is highly susceptible to side-channel attacks, unless constant-time 64-bit floating-point operations are used. This requirement is extremely platform-dependent, as noted in NIST's report. @@ -574,9 +574,9 @@ The number of tree layers in XMSS^MT provides a trade-off between signature size Due to the complexities described above, the XMSS and LMS are not a suitable replacement for classical signature schemes like RSA or ECDSA. Applications that expect a long lifetime of a signature, like firmware update or secure boot, are typical use cases where those schemes can be successfully applied. ### LMS scheme - key and signature sizes -The LMS scheme is characterized by four distinct parameter sets - underlying hash function (SHA2-256 or SHAKE-256), the length of the digest (24 or 32 bytes), LMS tree height - parameter that controls a maximal number of signatures that the private key can produce (possible values are 5,10,15,20,25) and the width of the Winternitz coefficients (see {{?RFC8554}}, section 4.1) that can be used to trade-off signing time for signature size (possible values are 1,2,4,8). Parameters can be mixed, providing 80 possible parametrizations of the scheme. +The LMS scheme is characterized by four distinct parameter sets - the underlying hash function (SHA2-256 or SHAKE-256), the length of the digest (24 or 32 bytes), LMS tree height - parameter that controls a maximal number of signatures that the private key can produce (possible values are 5,10,15,20,25), and the width of the Winternitz coefficients (see {{?RFC8554}}, section 4.1) that can be used to trade-off signing time for signature size (possible values are 1,2,4,8). Parameters can be mixed, providing 80 possible parametrizations of the scheme. -The public (PK) and private (SK) key size depends on the length of the digest (M). The signature size depends on the Winternitz parameter (W), the LMS tree height (H), and the length of the digest. The tables below provides key and signature sizes for parameterization with the digest size M=32 of the scheme. +The public (PK) and private (SK) key size depends on the length of the digest (M). The signature size depends on the Winternitz parameter (W), the LMS tree height (H), and the length of the digest. The table below provides key and signature sizes for parameterization with the digest size M=32 of the scheme. | PK | SK | W | H=5 | H=10 | H=15 | H=20 | H=25 | |----|----|---|------|------|------|------|------| @@ -589,14 +589,14 @@ The public (PK) and private (SK) key size depends on the length of the digest (M Within the hash-then-sign paradigm, the message is hashed before signing it. By pre-hashing, the onus of resistance to existential forgeries becomes heavily reliant on the collision-resistance of the hash function in use. The hash-then-sign paradigm has the ability to improve application performance by reducing the size of signed messages that need to be transmitted between application and cryptographic module, and making the signature size predictable and manageable. As a corollary, hashing remains mandatory even for short messages and assigns a further computational requirement onto the verifier. This makes the performance of hash-then-sign schemes more consistent, but not necessarily more efficient. Using a hash function to produce a fixed-size digest of a message ensures that the signature is compatible with a wide range of systems and protocols, regardless of the specific message size or format. Crucially for hardware security modules, Hash-then-Sign also significantly reduces the amount of data that needs to be transmitted and processed by a hardware security module. Consider scenarios such as a networked HSM located in a different data center from the calling application or a smart card connected over a USB interface. In these cases, streaming a message that is megabytes or gigabytes long can result in notable network latency, on-device signing delays, or even depletion of available on-device memory. -Note that the vast majority of Internet protocols that sign large messages already perform some level form of content hashing at the protocol level, so this tends to be more of a concern with proprietary cryptographic protocols, and protocols from non-IETF standards bodies. Protocols like TLS 1.3 and DNSSEC use the Hash-then-Sign paradigm. In TLS 1.3 {{?RFC8446}} CertificateVerify message, the content that is covered under the signature includes the transcript hash output (Section 4.4.1 of {{?RFC8446}}), while DNSSEC {{?RFC4033}} uses it to provide origin authentication and integrity assurance services for DNS data. Similarly, the Cryptographic Message Syntax (CMS) {{?RFC5652}} includes a mandatory message digest step before invoking the signature algorithm. +Note that the vast majority of Internet protocols that sign large messages already perform some form of content hashing at the protocol level, so this tends to be more of a concern with proprietary cryptographic protocols, and protocols from non-IETF standards bodies. Protocols like TLS 1.3 and DNSSEC use the Hash-then-Sign paradigm. In TLS 1.3 {{?RFC8446}} CertificateVerify messages, the content that is covered under the signature includes the transcript hash output (Section 4.4.1 of {{?RFC8446}}), while DNSSEC {{?RFC4033}} uses it to provide origin authentication and integrity assurance services for DNS data. Similarly, the Cryptographic Message Syntax (CMS) {{?RFC5652}} includes a mandatory message digest step before invoking the signature algorithm. -In the case of ML-DSA, it internally incorporates the necessary hash operations as part of its signing algorithm. ML-DSA directly takes the original message, applies a hash function internally, and then uses the resulting hash value for the signature generation process. In case of SLH-DSA, it internally performs randomized message compression using a keyed hash function that can process arbitrary length messages. In case of FN-DSA, a hash function is used as part of the signature process, it uses the SHAKE-256 hash function to derive a digest of the message being signed. Therefore, ML-DSA, FN-DSA, and SLH-DSA offer enhanced security over the traditional Hash-then-Sign paradigm because by incorporating dynamic key material into the message digest, a pre-computed hash collision on the message to be signed no longer yields a signature forgery. Applications requiring the performance and bandwidth benefits of Hash-then-Sign may still pre-hash at the protocol level prior to invoking ML-DSA, FN-DSA, or SLH-DSA, but protocol designers should be aware that doing so re-introduces the weakness that hash collisions directly yield signature forgeries. Signing the full un-digested message is strongly preferred where applications can tolerate it. +In the case of ML-DSA, it internally incorporates the necessary hash operations as part of its signing algorithm. ML-DSA directly takes the original message, applies a hash function internally, and then uses the resulting hash value for the signature generation process. In the case of SLH-DSA, it internally performs randomized message compression using a keyed hash function that can process arbitrary length messages. In the case of FN-DSA, the SHAKE-256 hash function is used as part of the signature process to derive a digest of the message being signed. Therefore, ML-DSA, FN-DSA, and SLH-DSA offer enhanced security over the traditional Hash-then-Sign paradigm because by incorporating dynamic key material into the message digest, a pre-computed hash collision on the message to be signed no longer yields a signature forgery. Applications requiring the performance and bandwidth benefits of Hash-then-Sign may still pre-hash at the protocol level prior to invoking ML-DSA, FN-DSA, or SLH-DSA, but protocol designers should be aware that doing so re-introduces the weakness that hash collisions directly yield signature forgeries. Signing the full un-digested message is strongly preferred where applications can tolerate it. # Recommendations for Security / Performance Tradeoffs {#RecSecurity} -The table below denotes the 5 security levels provided by NIST required for PQC algorithms. Neither NIST nor the IETF make any specific recommendations about which security level to use. In general, protocols will include algorithm choices at multiple levels so that users can choose the level appropriate to their policies and data classification, similar to how organizations today choose which size of RSA key to use. The security levels are defined as requiring computational resources comparable to or greater than an attack on AES (128, 192 and 256) and SHA2/SHA3 algorithms, i.e., exhaustive key recovery for AES and optimal collision search for SHA2/SHA3. This information is a re-print of information provided in the NIST PQC project {NIST} as of time of writing (July 2023). +The table below denotes the 5 security levels provided by NIST for PQC algorithms. Neither NIST nor the IETF make any specific recommendations about which security level to use. In general, protocols will include algorithm choices at multiple levels so that users can choose the level appropriate to their policies and data classification, similar to how organizations today choose which size of RSA key to use. The security levels are defined as requiring computational resources comparable to or greater than an attack on AES (128, 192 and 256) and SHA2/SHA3 algorithms, i.e., exhaustive key recovery for AES and optimal collision search for SHA2/SHA3. This information is a re-print of information provided in the NIST PQC project {{?NIST}} as of time of writing (July 2023). | PQ Security Level | AES/SHA(2/3) hardness | PQC Algorithm | | ----------------- | ----------------------------------------------- | ---------------------------------------------------------- | @@ -606,7 +606,7 @@ The table below denotes the 5 security levels provided by NIST required for PQC | 4 | SHA-384/SHA3-384 (collision search) | No algorithm tested at this level | | 5 | AES-256 (exhaustive key recovery) | ML-KEM-1024, FN-DSA-1024, ML-DSA-87, SLH-DSA-SHA2/SHAKE-256f/s | -Please note the SLH-DSA-x-yf/s "f/s" in the above table denotes whether its the SLH-DSA uses SHAKE or SHA-2 as an underlying hash function "x" and whether it is fast (f) version or small (s) version for "y" bit AES security level. Refer to {{?I-D.ietf-lamps-cms-sphincs-plus-02}} for further details on SLH-DSA algorithms. +Please note the SLH-DSA-x-yf/s "f/s" in the above table denotes whether its the SLH-DSA uses SHAKE or SHA-2 as an underlying hash function "x" and whether it is the fast (f) or small (s) version for "y" bit AES security level. Refer to {{?I-D.ietf-lamps-cms-sphincs-plus-02}} for further details on SLH-DSA algorithms. The following table discusses the signature size differences for similar SLH-DSA algorithm security levels with the "simple" version but for different categories i.e., (f) for fast verification and (s) for compactness/smaller. Both SHA-256 and SHAKE-256 parameterization output the same signature sizes, so both have been included. @@ -619,7 +619,7 @@ The following table discusses the signature size differences for similar SLH-DSA | 5 | SLH-DSA-{SHA2,SHAKE}-256f | 64 | 128 | 49856 | | 5 | SLH-DSA-{SHA2,SHAKE}-256s | 64 | 128 | 29792 | -The following table discusses the impact of performance on different security levels in terms of private key sizes, public key sizes and ciphertext/signature sizes. +The following table discusses the impact of performance on different security levels in terms of private key sizes, public key sizes, and ciphertext/signature sizes. | PQ Security Level | Algorithm | Public key size (in bytes) | Private key size (in bytes) | Ciphertext/Signature size (in bytes) | | ------------------ | -------------------------- | --------------------------- | --------------------------- | ------------------------------------ | @@ -636,7 +636,7 @@ The following table discusses the impact of performance on different security le In this section, we provide two tables for comparison of different KEMs and Signatures respectively, in the traditional and post-quantum scenarios. These tables will focus on the secret key sizes, public key sizes, and ciphertext/signature sizes for the PQC algorithms and their traditional counterparts of similar security levels. -The first table compares traditional vs. PQC KEMs in terms of security, public, private key sizes, and ciphertext sizes. +The first table compares traditional vs. PQC KEMs in terms of security, public and private key sizes, and ciphertext sizes. | PQ Security Level | Algorithm | Public key size (in bytes) | Private key size (in bytes) | Ciphertext size (in bytes) | | ----------------- | -------------------------- | --------------------------- | --------------------------- | ------------------------------------ | @@ -676,13 +676,13 @@ The PQ/T Hybrid Confidentiality property can be used to protect from a "Harvest 1. Concatenate hybrid key agreement scheme: The final shared secret that will be used as an input of the key derivation function is the result of the concatenation of the secrets established with each key agreement scheme. For example, in {{?I-D.ietf-tls-hybrid-design}}, the client uses the TLS supported groups extension to advertise support for a PQ/T hybrid scheme, and the server can select this group if it supports the scheme. The hybrid-aware client and server establish a hybrid secret by concatenating the two shared secrets, which is used as the shared secret in the existing TLS 1.3 key schedule. 2. Cascade hybrid key agreement scheme: The final shared secret is computed by applying as many iterations of the key derivation function as the number of key agreement schemes composing the hybrid key agreement scheme. For example, {{?RFC9370}} extends the Internet Key Exchange Protocol Version 2 (IKEv2) to allow one or more PQC algorithms in addition to the traditional algorithm to derive the final IKE SA keys using the cascade method as explained in Section 2.2.2 of {{?RFC9370}}. -Various instantiations of these two types of hybrid key agreement schemes have been explored. One must be careful when selecting which hybrid scheme to use. The chosen schemes at IETF are IND-CCA2 robust, that is IND-CCA2 security is guaranteed for the scheme as long as at least one of the component algorithms is IND-CCA2 secure. +Various instantiations of these two types of hybrid key agreement schemes have been explored. One must be careful when selecting which hybrid scheme to use. The chosen schemes at IETF are IND-CCA2 robust; that is, IND-CCA2 security is guaranteed for the scheme as long as at least one of the component algorithms is IND-CCA2 secure. ## PQ/T Hybrid Authentication The PQ/T Hybrid Authentication property can be utilized in scenarios where an on-path attacker possesses network devices equipped with CRQCs, capable of breaking traditional authentication protocols, or where an attacker can attack long-lived authenticated data such as CA certificates or signed software images. This property ensures authentication through a PQ/T hybrid scheme or a PQ/T hybrid protocol, as long as at least one component algorithm remains secure to provide the intended security level. For instance, a PQ/T hybrid certificate can be employed to facilitate a PQ/T hybrid authentication protocol. However, a PQ/T hybrid authentication protocol does not need to use a PQ/T hybrid certificate {{?I-D.ounsworth-pq-composite-keys}}; separate certificates could be used for individual component algorithms {{?I-D.ietf-lamps-cert-binding-for-multi-auth}}. -The frequency and duration of system upgrades and the time when CRQCs will become widely available need to be weighed in to determine whether and when to support the PQ/T Hybrid Authentication property. +The frequency and duration of system upgrades and the time when CRQCs will become widely available need to be weighed to determine whether and when to support the PQ/T Hybrid Authentication property. ## Additional Considerations @@ -690,15 +690,15 @@ It is also possible to use more than two algorithms together in a hybrid scheme, When combining keys in an "and" mode, it may make more sense to consider them to be a single composite key, instead of two keys. This generally requires fewer changes to various components of PKI ecosystems, many of which are not prepared to deal with two keys or dual signatures. To those protocol- or application-layer parsers, a "composite" algorithm composed of two "component" algorithms is simply a new algorithm, and support for adding new algorithms generally already exists. Treating multiple "component" keys as a single "composite" key also has security advantages such as preventing cross-protocol reuse of the individual component keys and guarantees about revoking or retiring all component keys together at the same time, especially if the composite is treated as a single object all the way down into the cryptographic module. All that needs to be done is to standardize the formats of how the two keys from the two algorithms are combined into a single data structure, and how the two resulting signatures or KEMs are combined into a single signature or KEM. The answer can be as simple as concatenation, if the lengths are fixed or easily determined. At time of writing, security research is ongoing as to the security properties of concatenation-based composite signatures and KEMs vs more sophisticated signature and KEM combiners, and in which protocol contexts those simpler combiners are sufficient. -One last consideration is the pairs of algorithms that can be combined. A recent trends in protocols is to only allow a small number of "known good" configurations that make sense, often referred to in cryptography as a "ciphersuite", instead of allowing arbitrary combinations of individual configuration choices that may interact in dangerous ways. The current consensus is that the same approach should be followed for combining cryptographic algorithms, and that "known good" pairs should be explicitly listed ("explicit composite"), instead of just allowing arbitrary combinations of any two crypto algorithms ("generic composite"). +One last consideration is the pairs of algorithms that can be combined. A recent trend in protocols is to only allow a small number of "known good" configurations that make sense, often referred to in cryptography as a "ciphersuite", instead of allowing arbitrary combinations of individual configuration choices that may interact in dangerous ways. The current consensus is that the same approach should be followed for combining cryptographic algorithms, and that "known good" pairs should be explicitly listed ("explicit composite"), instead of just allowing arbitrary combinations of any two crypto algorithms ("generic composite"). The same considerations apply when using multiple certificates to transport a pair of related keys for the same subject. Exactly how two certificates should be managed in order to avoid some of the pitfalls mentioned above is still an active area of investigation. Using two certificates keeps the certificate tooling simple and straightforward, but in the end simply moves the problems with requiring that both certs are intended to be used as a pair, must produce two signatures which must be carried separately, and both must validate, to the certificate management layer, where addressing these concerns in a robust way can be difficult. -An important security note when using particularly hybrid signature keys, but also to a lesser extent hybrid KEM keys, is key re-use. In traditional cryptography, problems can occur with so-called "cross-protocol attacks" when the same key can be used for multiple protocols; for example signing TLS handshakes and signing S/MIME emails. While it is not best-practice to re-use keys within the same protocol, for example using the same key for multiple S/MIME certificates for the same user, it is not generally catastrophic for security. However, key re-use becomes a large security problem within hybrids. Consider an \{RSA, ML-DSA\} hybrid key where the RSA key also appears within a single-algorithm certificate. In this case, an attacker could perform a "stripping attack" where they take some piece of data signed with the \{RSA, ML-DSA\} key, remove the ML-DSA signature and present the data as if it was intended for the RSA only certificate. This leads to a set of security definitions called "non-separability properties", which refers to how well the signature scheme resists various complexities of downgrade / stripping attacks {{?I-D.draft-ietf-pquip-hybrid-signature-spectrums}}. Therefore, implementers must either reuse the entire hybrid key as a whole, or perform fresh keygens of all component keys per usage, and must not take an existing key and reuse it as a component of a hybrid. +An important security note when using particularly hybrid signature keys, but also to a lesser extent hybrid KEM keys, is key re-use. In traditional cryptography, problems can occur with so-called "cross-protocol attacks" when the same key can be used for multiple protocols; for example, signing TLS handshakes and signing S/MIME emails. While it is not best-practice to re-use keys within the same protocol, for example, using the same key for multiple S/MIME certificates for the same user, it is not generally catastrophic for security. However, key re-use becomes a large security problem within hybrids. Consider an \{RSA, ML-DSA\} hybrid key where the RSA key also appears within a single-algorithm certificate. In this case, an attacker could perform a "stripping attack" where they take some piece of data signed with the \{RSA, ML-DSA\} key, remove the ML-DSA signature and present the data as if it was intended for the RSA only certificate. This leads to a set of security definitions called "non-separability properties", which refers to how well the signature scheme resists various complexities of downgrade / stripping attacks {{?I-D.draft-ietf-pquip-hybrid-signature-spectrums}}. Therefore, implementers must either reuse the entire hybrid key as a whole, or perform fresh keygens of all component keys per usage, and must not take an existing key and reuse it as a component of a hybrid. At least one scheme has been proposed that allows the pair of certificates to exist as a single certificate when being issued and managed, but dynamically split into individual certificates when needed (https://datatracker.ietf.org/doc/draft-bonnell-lamps-chameleon-certs/). -Another potential application of hybrids bears mentioning, even though it is not directly PQC-related. That is using hybrids to navigate inter-jurisdictional cryptographic connections. Traditional cryptography is already fragmented by jurisdiction, consider that while most jurisdictions support Elliptic Curve Diffie-Hellman, those in the United States will prefer the NIST curves while those in Germany will prefer the brainpool curves. China, Russia, and other jurisdictions have their own national cryptography standards. This situation of fragmented global cryptography standards is unlikely to improve with PQC. If "and" mode hybrids become standardized for the reasons mentioned above, then one could imagine leveraging them to create "ciphersuites" in which a single cryptographic operation simultaneously satisfies the cryptographic requirements of both endpoints. +Another potential application of hybrids bears mentioning, even though it is not directly PQC-related. That is using hybrids to navigate inter-jurisdictional cryptographic connections. Traditional cryptography is already fragmented by jurisdiction: consider that while most jurisdictions support Elliptic Curve Diffie-Hellman, those in the United States will prefer the NIST curves while those in Germany will prefer the brainpool curves. China, Russia, and other jurisdictions have their own national cryptography standards. This situation of fragmented global cryptography standards is unlikely to improve with PQC. If "and" mode hybrids become standardized for the reasons mentioned above, then one could imagine leveraging them to create "ciphersuites" in which a single cryptographic operation simultaneously satisfies the cryptographic requirements of both endpoints. Many of these points are still being actively explored and discussed, and the consensus may change over time. @@ -708,7 +708,7 @@ Many of these points are still being actively explored and discussed, and the co Classical cryptanalysis exploits weaknesses in algorithm design, mathematical vulnerabilities, or implementation flaws, that are exploitable with classical (i.e., non-quantum) hardware whereas quantum cryptanalysis harnesses the power of CRQCs to solve specific mathematical problems more efficiently. Another form of quantum cryptanalysis is 'quantum side-channel' attacks. In such attacks, a device under threat is directly connected to a quantum computer, which then injects entangled or superimposed data streams to exploit hardware that lacks protection against quantum side-channels. Both pose threats to the security of cryptographic algorithms, including those used in PQC. Developing and adopting new cryptographic algorithms resilient against these threats is crucial for ensuring long-term security in the face of advancing cryptanalysis techniques. -Recent attacks on the side-channel implementations using deep learning based power analysis have also shown that one needs to be cautious while implementing the required PQC algorithms in hardware. Two of the most recent works include: one attack on ML-KEM {{KyberSide}} and one attack on Saber {{SaberSide}}. Evolving threat landscape points to the fact that lattice based cryptography is indeed more vulnerable to side-channel attacks as in {{SideCh}}, {{LatticeSide}}. Consequently, there were some mitigation techniques for side channel attacks that have been proposed as in {{Mitigate1}}, {{Mitigate2}}, and {{Mitigate3}}. +Recent attacks on the side-channel implementations using deep learning based power analysis have also shown that one needs to be cautious while implementing the required PQC algorithms in hardware. Two of the most recent works include one attack on ML-KEM {{KyberSide}} and one attack on Saber {{SaberSide}}. An evolving threat landscape points to the fact that lattice based cryptography is indeed more vulnerable to side-channel attacks as in {{SideCh}}, {{LatticeSide}}. Consequently, there were some mitigation techniques for side channel attacks that have been proposed as in {{Mitigate1}}, {{Mitigate2}}, and {{Mitigate3}}. ## Cryptographic Agility