| Internet-Draft | Key Consistency and Discovery | March 2022 |
| Davidson, et al. | Expires 5 September 2022 | [Page] |
This document describes the key consistency and correctness requirements of protocols such as Privacy Pass, Oblivious DoH, and Oblivious HTTP for user privacy. It discusses several mechanisms and proposals for enabling user privacy in varying threat models. In concludes with discussion of open problems in this area.¶
This note is to be removed before publishing as an RFC.¶
Discussion of this document takes place on the mailing list (), which is archived at .¶
Source for this draft and an issue tracker can be found at https://github.com/chris-wood/key-consitency.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 5 September 2022.¶
Copyright (c) 2022 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
Several proposed privacy-enhancing protocols such as Privacy Pass [PRIVACY-PASS], Oblivious DoH [ODOH], and Oblivious HTTP [OHTTP] require clients to obtain and use a public key for execution. For example, Privacy Pass public keys are used by clients for validating privately issued tokens for anonymous session resumption. Oblivious DoH and HTTP both use public keys to encrypt messages to a particular server.¶
User privacy in these systems depends on users receiving a key that many, if not all, other users receive. If a user were to receive a public key that was specific to them, or restricted to a small set of users, then use of that public key could be used to learn targeted information about the user. Users also need to receive the correct public key.¶
In this document, we elaborate on these core requirements, and survey various system designs that might be used to satisfy them. The purpose of this document is to highlight challenges in building and deploying solutions to this problem.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
This document defines the following terms:¶
A mechanism for providing clients with a consistent view of cryptographic key material within a period of time.¶
A system that embeds one or more key consistency and correctness systems.¶
The KCCS's consistency model is dependent on the implementation and reliant system's threat model.¶
Privacy-focused protocols which rely on widely shared public keys typically require keys be consistent and correct. Informally, key consistency is the requirement that all users who communicate with an entity share the same view of the key associated with that entity; key correctness is that the key's secret information is controlled by the intended entity and is not known to be available to an external attacker.¶
Some protocols depend on large sets of users with consistent keys for privacy reasons. Specifically, all users with a consistent key represent an anonymity set wherein each user of the key in that set is indistinguishable from the rest. An attacker that can actively cause inconsistent views of keys can therefore compromise user privacy.¶
An attacker that can cause a user to use an incorrect key will likely compromise the entire protocol, not just privacy.¶
Reliant systems must also consider agility when trying to satisfy these requirements. A naive solution to ensuring consistent and correct keys is to only use a single, fixed key pair for the entirety of the system. Users can then embed this key into software or elsewhere as needed, without any additional mechanics or controls to ensure that other users have a different key. However, this solution clearly is not viable in practice. If the corresponding key is compromised, the system fails. Rotation must therefore be supported, and in doing so, users need some mechanism to ensure that newly rotated keys are consistent and correct.¶
Operationally, servers rotating keys may likely need to accommodate distributed system state-synchronization issues without sacrificing availability. Some systems and protocols may choose to prioritize strong consistency over availability, but this document assumes that availability is preferred to total consistency.¶
There are a variety of ways in which reliant systems may build key consistency and correct systems (KCCS), ranging in operational complexity to ease-of-implementation. In this section, we survey a number of possible solutions. The viability of each varies depending on the applicable threat model, external dependencies, and overall reliant system's requirements.¶
We do not include the fixed public key model from Section 3, as this is likely not a viable solution for systems and protocols in practice. In all scenarios, the server corresponding to the desired key is considered malicious.¶
In this model, users would directly query servers for their corresponding public key, as shown below.¶
+----------+ +----------+ | | | | | Client +--------------> Server | | | | | +----------+ +----------+
The properties of this solution depend on external mechanisms in place to ensure consistency or correctness. Absent any such mechanisms, servers can produce unique keys for users without detection. External mechanisms to ensure consistency here might include, though are not limited to:¶
The precise external mechanism used here depends largely on the threat model. If there is a trusted external log for keys, this may be a viable solution.¶
In this model, there exists a proxy that fetches keys from servers on behalf of multiple users, as shown below.¶
+----------+
| |
| Client +----------+
| | |
+----------+ |
|
+----------+ +\/--------+ +----------+
| | | | | |
| Client +---------> Proxy +------> Server |
| | | | | |
+----------+ +^---------+ +----------+
x |
x |
+----------+ |
| | |
| Client +----------+
| |
+----------+
If this proxy is trusted, then all users which request a key from this server are assured they have a consistent view of the server key. However, if this proxy is not trusted, operational risks may arise:¶
Mitigating these risks may require tamper-proof logs as in Section 4.1, or via user gossip protocols.¶
In this model, users leverage multiple, non-colluding proxies to fetch keys from servers, as shown below.¶
+----------+
| |
+---------------> Proxy +-----------+
| | | |
| +----------+ |
| |
+----------+ +----------+ +----\/----+
| │ | | | |
| Client +---------> Proxy +------> Server |
| │ | | | |
+----------+ +----------+ +----^-----+
| x |
| x |
| +----------+ |
| | | |
+--------------> Proxy +-----------+
| |
+----------+
These proxies are ideally spread across multiple vantage points. Examples of proxies include anonymous systems such as Tor. Tor proxies are general purpose and operate at a lower layer, on arbitrary communication flows, and therefore they are oblivious to clients fetching keys. A large set of untrusted proxies that are aware of key fetch requests (Section 4.2) may be used in a similar way. Depending on how clients fetch such keys from servers, it may become more difficult for servers to uniquely target individual users with unique keys without detection. This is especially true as the number of users of these anonymity networks increases. However, beyond Tor, there does not exist a special-purpose anonymity network for this purpose.¶
In this model, servers publish keys in an external database and clients fetch keys from the database, as shown below.¶
+----------+
| |
| Client +-----------+
| | |
+----------+ |
|
+----------+ +----------+ +----------+
| | | | | |
| Client +---------> Database <------+ Server |
| | | | | |
+----------+ +----------+ +----------+
x |
x |
+----------+ |
| | |
| Client +-----------+
| |
+----------+
The database is expected to have a table that asserts mappings between server names and keys. Examples of such databases are as follows:¶
For privacy, users should either download the entire database and query it locally, or remotely query the database using privacy-preserving queries (e.g., a private information retrieval (PIR) protocol). In the case where the database is downloaded locally, it should be considered stale and re-fetched periodically. The frequency of such updates can likely be infrequent in practice, as frequent key updates or rotations may affect privacy; see Section 5 for details. Downloading the entire database works best if there are a small number of entries, as it does not otherwise impose bandwidth costs on each client that may be impractical.¶
In addition to ensuring that there is one key at any time, or a limited number keys, any system needs to ensure that a server cannot rotate its keys too often in order to divide clients into smaller groups based on when keys are acquired. Such considerations are already highlighted within the Privacy Pass ecosystem, more discussion can be found at [PRIVACY-PASS-ARCH]. Setting a minimum validity period limits the ability of a server to rotate keys, but also limits the rate of key rotation.¶
The other schemes described here all attempt to directly limit the number of keys that a client might accept. However, by changing how keys are used, clients can impose costs on servers that might discourage key diversity.¶
Protocols that have distinctly separate processes for acquiring and using keys might benefit from moving consistency checks to the usage part of the protocol. Correctness might be guaranteed through a relatively simple process, such obtaining keys directly from a server. A separate correctness check is then applied before keys are used.¶
Anonymous queries to verify key consistency can be used prior to use of keys. A request for the current key (or limited set of keys) will reveal if the key that was acquired is different than the original. If the key that was originally obtained is not included, the client can abort any use of the key.¶
It is important that any validation process not carry any information that might tie it to the original key discovery process or that the system providing verification be trusted. A proxy (see Section 4.2) might be sufficient for providing anonymity, though more robust anonymity protections (see Section 4.3) could provide stronger guarantees. Querying a database (see Section 4.4) might provide independent verification if that database can be trusted not to provide answers that change based on client identity.¶
Key-based encryption has a client encrypt the information that it sends to a server, such as a token or signed object generated with the server keys. This encryption uses a key derived from the key configuration, specifically not including any form of key identifier along with the encrypted information. If key derivation for the encryption uses a pre-image resistant function (like HKDF), the server can only decrypt the information if it knows the key configuration. As there is no information the server can use to identify which key was used, it is forced to perform trial decryption if it wants to use multiple keys.¶
These costs are only linear in terms of the number of active keys. This doesn't prevent the use of multiple keys, it only makes their use incrementally more expensive. Trial decryption costs can be increased by choosing a time- or memory-hard function such as [ARGON2] to generate keys.¶
Encrypting this way could provide better latency properties than a separate check.¶
The model in Section 4.3 seems to be the most lightweight and easy-to-deploy mechanism for ensuring key consistency and correctness. However, it remains unclear if there exists such an anonymity network that can scale to the widespread adoption of and requirements of protocols like Privacy Pass, Oblivious DoH, or Oblivious HTTP. Existing infrastructure based on technologies like Certificate Transparency or Key Transparency may work, but there is currently no general purpose system for transparency of opaque keys (or other application data).¶
This document discusses several models that systems might use to implement public key discovery while ensuring key consistency and correctness. It does not make any recommendations for such models as the best model depends on differing operational requirements and threat models.¶