Internet-Draft Oblivious HTTP February 2022
Thomson & Wood Expires 19 August 2022 [Page]
Workgroup:
Oblivious HTTP Application Intermediation
Internet-Draft:
draft-ietf-ohai-ohttp-01
Published:
Intended Status:
Standards Track
Expires:
Authors:
M. Thomson
Mozilla
C.A. Wood
Cloudflare

Oblivious HTTP

Abstract

This document describes a system for the forwarding of encrypted HTTP messages. This allows a client to make multiple requests of a server without the server being able to link those requests to the client or to identify the requests as having come from the same client.

About This Document

This note is to be removed before publishing as an RFC.

The latest revision of this draft can be found at https://ietf-wg-ohai.github.io/oblivious-http/draft-ietf-ohai-ohttp.html. Status information for this document may be found at https://datatracker.ietf.org/doc/draft-ietf-ohai-ohttp/.

Discussion of this document takes place on the Oblivious HTTP Application Intermediation Working Group mailing list (mailto:ohai@ietf.org), which is archived at https://mailarchive.ietf.org/arch/browse/ohai/.

Source for this draft and an issue tracker can be found at https://github.com/ietf-wg-ohai/oblivious-http.

Status of This Memo

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 19 August 2022.

Table of Contents

1. Introduction

The act of making a request using HTTP reveals information about the client identity to a server. Though the content of requests might reveal information, that is information under the control of the client. In comparison, the source address on the connection reveals information that a client has only limited control over.

Even where an IP address is not directly attributed to an individual, the use of an address over time can be used to correlate requests. Servers are able to use this information to assemble profiles of client behavior, from which they can make inferences about the people involved. The use of persistent connections to make multiple requests improves performance, but provides servers with additional certainty about the identity of clients in a similar fashion.

Use of an HTTP proxy can provide a degree of protection against servers correlating requests. Systems like virtual private networks or the Tor network [Dingledine2004], provide other options for clients.

Though the overhead imposed by these methods varies, the cost for each request is significant. Preventing request linkability requires that each request use a completely new TLS connection to the server. At a minimum, this requires an additional round trip to the server in addition to that required by the request. In addition to having high latency, there are significant secondary costs, both in terms of the number of additional bytes exchanged and the CPU cost of cryptographic computations.

This document describes a method of encapsulation for binary HTTP messages [BINARY] using Hybrid Public Key Encryption (HPKE; [HPKE]). This protects the content of both requests and responses and enables a deployment architecture that can separate the identity of a requester from the request.

Though this scheme requires that servers and proxies explicitly support it, this design represents a performance improvement over options that perform just one request in each connection. With limited trust placed in the proxy (see Section 8), clients are assured that requests are not uniquely attributed to them or linked to other requests.

2. Conventions and Definitions

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.

Encapsulated Request:

An HTTP request that is encapsulated in an HPKE-encrypted message; see Section 5.1.

Encapsulated Response:

An HTTP response that is encapsulated in an HPKE-encrypted message; see Section 5.2.

Oblivious Proxy Resource:

An intermediary that forwards requests and responses between clients and a single oblivious request resource.

Oblivious Request Resource:

A resource that can receive an encapsulated request, extract the contents of that request, forward it to an oblivious target resource, receive a response, encapsulate that response, then return that response.

Oblivious Target Resource:

The resource that is the target of an encapsulated request. This resource logically handles only regular HTTP requests and responses and so might be ignorant of the use of oblivious HTTP to reach it.

This draft includes pseudocode that uses the functions and conventions defined in [HPKE].

Encoding an integer to a sequence of bytes in network byte order is described using the function encode(n, v), where n is the number of bytes and v is the integer value. The function len() returns the length of a sequence of bytes.

Formats are described using notation from Section 1.3 of [QUIC].

3. Overview

A client learns the following:

This information allows the client to make a request of an oblivious target resource without that resource having only a limited ability to correlate that request with the client IP or other requests that the client might make to that server.

Client Proxy Request Target Resource Resource Resource Encapsulated Request Encapsulated Request Request Response Encapsulated Response Encapsulated Response
Figure 1: Overview of Oblivious HTTP

In order to make a request to an oblivious target resource, the following steps occur, as shown in Figure 1:

  1. The client constructs an HTTP request for an oblivious target resource.
  2. The client encodes the HTTP request in a binary HTTP message and then encapsulates that message using HPKE and the process from Section 5.1.
  3. The client sends a POST request to the oblivious proxy resource with the encapsulated request as the content of that message.
  4. The oblivious proxy resource forwards this request to the oblivious request resource.
  5. The oblivious request resource receives this request and removes the HPKE protection to obtain an HTTP request.
  6. The oblivious request resource makes an HTTP request that includes the target URI, method, fields, and content of the request it acquires.
  7. The oblivious target resource answers this HTTP request with an HTTP response.
  8. The oblivious request resource encapsulates the HTTP response following the process in Section 5.2 and sends this in response to the request from the oblivious proxy resource.
  9. The oblivious proxy resource forwards this response to the client.
  10. The client removes the encapsulation to obtain the response to the original request.

3.1. Applicability

Oblivious HTTP has limited applicability. Many uses of HTTP benefit from being able to carry state between requests, such as with cookies ([RFC6265]), authentication (Section 11 of [HTTP]), or even alternative services ([RFC7838]). Oblivious HTTP seeks to prevent this sort of linkage, which requires that applications not carry state between requests.

Oblivious HTTP is primarily useful where privacy risks associated with possible stateful treatment of requests are sufficiently negative that the cost of deploying this protocol can be justified. Oblivious HTTP is simpler and less costly than more robust systems, like Prio ([PRIO]) or Tor ([Dingledine2004]), which can provide stronger guarantees at higher operational costs.

Oblivious HTTP is more costly than a direct connection to a server. Some costs, like those involved with connection setup, can be amortized, but there are several ways in which oblivious HTTP is more expensive than a direct request:

  • Each oblivious request requires at least two regular HTTP requests, which adds latency.
  • Each request is expanded in size with additional HTTP fields, encryption-related metadata, and AEAD expansion.
  • Deriving cryptographic keys and applying them for request and response protection takes non-negligible computational resources.

Examples of where preventing the linking of requests might justify these costs include:

  • DNS queries. DNS queries made to a recursive resolver reveal information about the requester, particularly if linked to other queries.
  • Telemetry submission. Applications that submit reports about their usage to their developers might use oblivious HTTP for some types of moderately sensitive data.

These are examples of requests where there is information in a request that - if it were connected to the identity of the user - might allow a server to learn something about that user even if the identity of the user is pseudonymous. Other examples include the submission of anonymous surveys, making search queries, or requesting location-specific content (such as retrieving tiles of a map display).

4. Key Configuration

A client needs to acquire information about the key configuration of the oblivious request resource in order to send encapsulated requests.

In order to ensure that clients do not encapsulate messages that other entities can intercept, the key configuration MUST be authenticated and have integrity protection.

This document describes the "application/ohttp-keys" media type; see Section 4.2. This media type might be used, for example with HTTPS, as part of a system for configuring or discovering key configurations. Note however that such a system needs to consider the potential for key configuration to be used to compromise client privacy; see Section 9.

Specifying a format for expressing the information a client needs to construct an encapsulated request ensures that different client implementations can be configured in the same way. This also enables advertising key configurations in a consistent format.

A client might have multiple key configurations to select from when encapsulating a request. Clients are responsible for selecting a preferred key configuration from those it supports. Clients need to consider both the key encapsulation method (KEM) and the combinations of key derivation function (KDF) and authenticated encryption with associated data (AEAD) in this decision.

Evolution of the key configuration format is supported through the definition of new formats that are identified by new media types.

4.1. Key Configuration Encoding

A single key configuration consists of a key identifier, a public key, an identifier for the KEM that the public key uses, and a set HPKE symmetric algorithms. Each symmetric algorithm consists of an identifier for a KDF and an identifier for an AEAD.

Figure 2 shows a single key configuration, KeyConfig, that is expressed using the TLS syntax; see Section 3 of [TLS].

opaque HpkePublicKey[Npk];
uint16 HpkeKemId;
uint16 HpkeKdfId;
uint16 HpkeAeadId;

struct {
  HpkeKdfId kdf_id;
  HpkeAeadId aead_id;
} HpkeSymmetricAlgorithms;

struct {
  uint8 key_id;
  HpkeKemId kem_id;
  HpkePublicKey public_key;
  HpkeSymmetricAlgorithms cipher_suites<4..2^16-4>;
} KeyConfig;
Figure 2: A Single Key Configuration

The types HpkeKemId, HpkeKdfId, and HpkeAeadId identify a KEM, KDF, and AEAD respectively. The definitions for these identifiers and the semantics of the algorithms they identify can be found in [HPKE]. The Npk parameter corresponding to the HpkeKdfId can be found in [HPKE].

4.2. Key Configuration Media Type

The "application/ohttp-keys" format is a media type that identifies a serialized collection of key configurations. The content of this media type comprises one or more key configuration encodings (see Section 4.1) that are concatenated.

Type name:

application

Subtype name:

ohttp-keys

Required parameters:

N/A

Optional parameters:

None

Encoding considerations:

only "8bit" or "binary" is permitted

Security considerations:

see Section 8

Interoperability considerations:

N/A

Published specification:

this specification

Applications that use this media type:

N/A

Fragment identifier considerations:

N/A

Additional information:
Magic number(s):
N/A
Deprecated alias names for this type:
N/A
File extension(s):
N/A
Macintosh file type code(s):
N/A
Person and email address to contact for further information:

see Authors' Addresses section

Intended usage:

COMMON

Restrictions on usage:

N/A

Author:

see Authors' Addresses section

Change controller:

IESG

5. HPKE Encapsulation

HTTP message encapsulation uses HPKE for request and response encryption.

An encapsulated HTTP request contains a binary-encoded HTTP message [BINARY] and no other fields; see Figure 3.

Request {
  Binary HTTP Message (..),
}
Figure 3: Plaintext Request Content

An Encapsulated Request is comprised of a length-prefixed key identifier and a HPKE-protected request message. HPKE protection includes an encapsulated KEM shared secret (or enc), plus the AEAD-protected request message. An Encapsulated Request is shown in Figure 4. Section 5.1 describes the process for constructing and processing an Encapsulated Request.

Encapsulated Request {
  Key Identifier (8),
  KEM Identifier (16),
  KDF Identifier (16),
  AEAD Identifier (16),
  Encapsulated KEM Shared Secret (8*Nenc),
  AEAD-Protected Request (..),
}
Figure 4: Encapsulated Request

The Nenc parameter corresponding to the HpkeKdfId can be found in Section 7.1 of [HPKE].

An encapsulated HTTP response includes a binary-encoded HTTP message [BINARY] and no other content; see Figure 5.

Response {
  Binary HTTP Message (..),
}
Figure 5: Plaintext Response Content

Responses are bound to responses and so consist only of AEAD-protected content. Section 5.2 describes the process for constructing and processing an Encapsulated Response.

Encapsulated Response {
  Nonce (Nk),
  AEAD-Protected Response (..),
}
Figure 6: Encapsulated Response

The Nenc and Nk parameters corresponding to the HpkeKdfId can be found in [HPKE]. Nenc refers to the size of the encapsulated KEM shared secret, in bytes; Nk refers to the size of the AEAD key for the HPKE ciphersuite, in bits.

5.1. Encapsulation of Requests

Clients encapsulate a request request using values from a key configuration:

  • the key identifier from the configuration, keyID, with the corresponding KEM identified by kemID,
  • the public key from the configuration, pkR, and
  • a selected combination of KDF, identified by kdfID, and AEAD, identified by aeadID.

The client then constructs an encapsulated request, enc_request, from a binary encoded HTTP request, request, as follows:

  1. Compute an HPKE context using pkR and a label of "message/bhttp request", yielding context and encapsulation key enc.
  2. Construct associated data, aad, by concatenating the values of keyID, kemID, kdfID, and aeadID, as one 8-bit integer and three 16-bit integers, respectively, each in network byte order.
  3. Encrypt (seal) request with aad as associated data using context, yielding ciphertext ct.
  4. Concatenate the values of aad, enc, and ct, yielding an Encapsulated Request enc_request.

Note that enc is of fixed-length, so there is no ambiguity in parsing this structure.

In pseudocode, this procedure is as follows:

enc, context = SetupBaseS(pkR, "message/bhttp request")
aad = concat(encode(1, keyID),
             encode(2, kemID),
             encode(2, kdfID),
             encode(2, aeadID))
ct = context.Seal(aad, request)
enc_request = concat(aad, enc, ct)

Servers decrypt an Encapsulated Request by reversing this process. Given an Encapsulated Request enc_request, a server:

  1. Parses enc_request into keyID, kemID, kdfID, aeadID, enc, and ct (indicated using the function parse() in pseudocode). The server is then able to find the HPKE private key, skR, corresponding to keyID.

    a. If keyID does not identify a key matching the type of kemID, the server returns an error.

    b. If kdfID and aeadID identify a combination of KDF and AEAD that the server is unwilling to use with skR, the server returns an error.

  2. Compute an HPKE context using skR, a label of "message/bhttp request", and the encapsulated key enc, yielding context.
  3. Construct additional associated data, aad, from keyID, kemID, kdfID, and aeadID or as the first seven bytes of enc_request.
  4. Decrypt ct using aad as associated data, yielding request or an error on failure. If decryption fails, the server returns an error.

In pseudocode, this procedure is as follows:

keyID, kemID, kdfID, aeadID, enc, ct = parse(enc_request)
aad = concat(encode(1, keyID),
             encode(2, kemID),
             encode(2, kdfID),
             encode(2, aeadID))
context = SetupBaseR(enc, skR, "message/bhttp request")
request, error = context.Open(aad, ct)

5.2. Encapsulation of Responses

Given an HPKE context context, a request message request, and a response response, servers generate an Encapsulated Response enc_response as follows:

  1. Export a secret secret from context, using the string "message/bhttp response" as context. The length of this secret is max(Nn, Nk), where Nn and Nk are the length of AEAD key and nonce associated with context.
  2. Generate a random value of length max(Nn, Nk) bytes, called response_nonce.
  3. Extract a pseudorandom key prk using the Extract function provided by the KDF algorithm associated with context. The ikm input to this function is secret; the salt input is the concatenation of enc (from enc_request) and response_nonce
  4. Use the Expand function provided by the same KDF to extract an AEAD key key, of length Nk - the length of the keys used by the AEAD associated with context. Generating key uses a label of "key".
  5. Use the same Expand function to extract a nonce nonce of length Nn - the length of the nonce used by the AEAD. Generating nonce uses a label of "nonce".
  6. Encrypt response, passing the AEAD function Seal the values of key, nonce, empty aad, and a pt input of request, which yields ct.
  7. Concatenate response_nonce and ct, yielding an Encapsulated Response enc_response. Note that response_nonce is of fixed-length, so there is no ambiguity in parsing either response_nonce or ct.

In pseudocode, this procedure is as follows:

secret = context.Export("message/bhttp response", Nk)
response_nonce = random(max(Nn, Nk))
salt = concat(enc, response_nonce)
prk = Extract(salt, secret)
aead_key = Expand(prk, "key", Nk)
aead_nonce = Expand(prk, "nonce", Nn)
ct = Seal(aead_key, aead_nonce, "", response)
enc_response = concat(response_nonce, ct)

Clients decrypt an Encapsulated Request by reversing this process. That is, they first parse enc_response into response_nonce and ct. They then follow the same process to derive values for aead_key and aead_nonce.

The client uses these values to decrypt ct using the Open function provided by the AEAD. Decrypting might produce an error, as follows:

reponse, error = Open(aead_key, aead_nonce, "", ct)

6. HTTP Usage

A client interacts with the oblivious proxy resource by constructing an encapsulated request. This encapsulated request is included as the content of a POST request to the oblivious proxy resource. This request MUST only contain those fields necessary to carry the encapsulated request: a method of POST, a target URI of the oblivious proxy resource, a header field containing the content type (see (Section 7), and the encapsulated request as the request content. Clients MAY include fields that do not reveal information about the content of the request, such as Alt-Used [ALT-SVC], or information that it trusts the oblivious proxy resource to remove, such as fields that are listed in the Connection header field.

The oblivious proxy resource interacts with the oblivious request resource by constructing a request using the same restrictions as the client request, except that the target URI is the oblivious request resource. The content of this request is copied from the client. The oblivious proxy resource MUST NOT add information about the client to this request.

When a response is received from the oblivious request resource, the oblivious proxy resource forwards the response according to the rules of an HTTP proxy; see Section 7.6 of [HTTP].

An oblivious request resource, if it receives any response from the oblivious target resource, sends a single 200 response containing the encapsulated response. Like the request from the client, this response MUST only contain those fields necessary to carry the encapsulated response: a 200 status code, a header field indicating the content type, and the encapsulated response as the response content. As with requests, additional fields MAY be used to convey information that does not reveal information about the encapsulated response.

An oblivious request resource acts as a gateway for requests to the oblivious target resource (see Section 7.6 of [HTTP]). The one exception is that any information it might forward in a response MUST be encapsulated, unless it is responding to errors it detects before removing encapsulation of the request; see Section 6.2.

6.1. Informational Responses

This encapsulation does not permit progressive processing of responses. Though the binary HTTP response format does support the inclusion of informational (1xx) status codes, the AEAD encapsulation cannot be removed until the entire message is received.

In particular, the Expect header field with 100-continue (see Section 10.1.1 of [HTTP]) cannot be used. Clients MUST NOT construct a request that includes a 100-continue expectation; the oblivious request resource MUST generate an error if a 100-continue expectation is received.

6.2. Errors

A server that receives an invalid message for any reason MUST generate an HTTP response with a 4xx status code.

Errors detected by the oblivious proxy resource and errors detected by the oblivious request resource before removing protection (including being unable to remove encapsulation for any reason) result in the status code being sent without protection in response to the POST request made to that resource.

Errors detected by the oblivious request resource after successfully removing encapsulation and errors detected by the oblivious target resource MUST be sent in an encapsulated response.

7. Media Types

Media types are used to identify encapsulated requests and responses.

Evolution of the format of encapsulated requests and responses is supported through the definition of new formats that are identified by new media types.

7.1. message/ohttp-req Media Type

The "message/ohttp-req" identifies an encapsulated binary HTTP request. This is a binary format that is defined in Section 5.1.

Type name:

message

Subtype name:

ohttp-req

Required parameters:

N/A

Optional parameters:

None

Encoding considerations:

only "8bit" or "binary" is permitted

Security considerations:

see Section 8

Interoperability considerations:

N/A

Published specification:

this specification

Applications that use this media type:

N/A

Fragment identifier considerations:

N/A

Additional information:
Magic number(s):
N/A
Deprecated alias names for this type:
N/A
File extension(s):
N/A
Macintosh file type code(s):
N/A
Person and email address to contact for further information:

see Authors' Addresses section

Intended usage:

COMMON

Restrictions on usage:

N/A

Author:

see Authors' Addresses section

Change controller:

IESG

7.2. message/ohttp-res Media Type

The "message/ohttp-res" identifies an encapsulated binary HTTP response. This is a binary format that is defined in Section 5.2.

Type name:

message

Subtype name:

ohttp-res

Required parameters:

N/A

Optional parameters:

None

Encoding considerations:

only "8bit" or "binary" is permitted

Security considerations:

see Section 8

Interoperability considerations:

N/A

Published specification:

this specification

Applications that use this media type:

N/A

Fragment identifier considerations:

N/A

Additional information:
Magic number(s):
N/A
Deprecated alias names for this type:
N/A
File extension(s):
N/A
Macintosh file type code(s):
N/A
Person and email address to contact for further information:

see Authors' Addresses section

Intended usage:

COMMON

Restrictions on usage:

N/A

Author:

see Authors' Addresses section

Change controller:

IESG

8. Security Considerations

In this design, a client wishes to make a request of a server that is authoritative for the oblivious target resource. The client wishes to make this request without linking that request with either:

  1. The identity at the network and transport layer of the client (that is, the client IP address and TCP or UDP port number the client uses to create a connection).
  2. Any other request the client might have made in the past or might make in the future.

In order to ensure this, the client selects a proxy (that serves the oblivious proxy resource) that it trusts will protect this information by forwarding the encapsulated request and response without passing the server (that serves the oblivious request resource).

In this section, a deployment where there are three entities is considered:

To achieve the stated privacy goals, the oblivious proxy resource cannot be operated by the same entity as the oblivious request resource. However, colocation of the oblivious request resource and oblivious target resource simplifies the interactions between those resources without affecting client privacy.

As a consequence of this configuration, Oblivious HTTP prevents linkability described above. Informally, this means:

  1. Requests and responses are known only to clients and targets in possession of the corresponding response encapsulation key and HPKE keying material. In particular, the oblivious proxy knows the origin and destination of an encapsulated request and response, yet does not know the decapsulated contents. Likewise, targets know only the oblivious request origin, i.e., the proxy, and the decapsulated request. Only the client knows both the plaintext request and response.
  2. Targets cannot link requests from the same client in the absence of unique per-client keys.

Traffic analysis that might affect these properties are outside the scope of this document; see Section 8.2.2.

A formal analysis of Oblivious HTTP is in [OHTTP-ANALYSIS].

8.1. Client Responsibilities

Clients MUST ensure that the key configuration they select for generating encapsulated requests is integrity protected and authenticated so that it can be attributed to the oblivious request resource; see Section 4.

Since clients connect directly to the proxy instead of the target, application configurations wherein clients make policy decisions about target connections, e.g., to apply certificate pinning, are incompatible with Oblivious HTTP. In such cases, alternative technologies such as HTTP CONNECT (Section 9.3.6 of [HTTP]) can be used. Applications could implement related policies on key configurations and proxy connections, though these might not provide the same properties as policies enforced directly on target connections. When this difference is relevant, applications can instead connect directly to the target at the cost of either privacy or performance.

Clients MUST NOT include identifying information in the request that is encapsulated. Identifying information includes cookies [COOKIES], authentication credentials or tokens, and any information that might reveal client-specific information such as account credentials.

Clients cannot carry connection-level state between requests as they only establish direct connections to the proxy responsible for the oblivious proxy resource. However, clients need to ensure that they construct requests without any information gained from previous requests. Otherwise, the server might be able to use that information to link requests. Cookies [COOKIES] are the most obvious feature that MUST NOT be used by clients. However, clients need to include all information learned from requests, which could include the identity of resources.

Clients MUST generate a new HPKE context for every request, using a good source of entropy ([RANDOM]) for generating keys. Key reuse not only risks requests being linked, reuse could expose request and response contents to the proxy.

The request the client sends to the oblivious proxy resource only requires minimal information; see Section 6. The request that carries the encapsulated request and is sent to the oblivious proxy resource MUST NOT include identifying information unless the client ensures that this information is removed by the proxy. A client MAY include information only for the oblivious proxy resource in header fields identified by the Connection header field if it trusts the proxy to remove these as required by Section 7.6.1 of [HTTP]. The client needs to trust that the proxy does not replicate the source addressing information in the request it forwards.

Clients rely on the oblivious proxy resource to forward encapsulated requests and responses. However, the proxy can only refuse to forward messages, it cannot inspect or modify the contents of encapsulated requests or responses.

8.2. Proxy Responsibilities

The proxy that serves the oblivious proxy resource has a very simple function to perform. For each request it receives, it makes a request of the oblivious request resource that includes the same content. When it receives a response, it sends a response to the client that includes the content of the response from the oblivious request resource. When generating a request, the proxy MUST follow the forwarding rules in Section 7.6 of [HTTP].

A proxy can also generate responses, though it assumed to not be able to examine the content of a request (other than to observe the choice of key identifier, KDF, and AEAD), so it is also assumed that it cannot generate an encapsulated response.

A proxy MUST NOT add information about the client identity when forwarding requests. This includes the Via field, the Forwarded field [FORWARDED], and any similar information. A client does not depend on the proxy using an authenticated and encrypted connection to the oblivious request resource, only that information about the client not be attached to forwarded requests.

8.2.1. Denial of Service

As there are privacy benefits from having a large rate of requests forwarded by the same proxy (see Section 8.2.2), servers that operate the oblivious request resource might need an arrangement with proxies. This arrangement might be necessary to prevent having the large volume of requests being classified as an attack by the server.

If a server accepts a larger volume of requests from a proxy, it needs to trust that the proxy does not allow abusive levels of request volumes from clients. That is, if a server allows requests from the proxy to be exempt from rate limits, the server might want to ensure that the proxy applies a rate limiting policy that is acceptable to the server.

Servers that enter into an agreement with a proxy that enables a higher request rate might choose to authenticate the proxy to enable the higher rate.

8.2.2. Linkability Through Traffic Analysis

As the time at which encapsulated request or response messages are sent can reveal information to a network observer. Though messages exchanged between the oblivious proxy resource and the oblivious request resource might be sent in a single connection, traffic analysis could be used to match messages that are forwarded by the proxy.

A proxy could, as part of its function, add delays in order to increase the anonymity set into which each message is attributed. This could latency to the overall time clients take to receive a response, which might not be what some clients want.

A proxy can use padding to reduce the effectiveness of traffic analysis. Padding is a capability provided by binary HTTP messages; see Section 3.8 of [BINARY].

A proxy that forwards large volumes of exchanges can provide better privacy by providing larger sets of messages that need to be matched.

8.3. Server Responsibilities

A server that operates both oblivious request and oblivious target resources is responsible for removing request encapsulation, generating a response the encapsulated request, and encapsulating the response.

Servers should account for traffic analysis based on response size or generation time. Techniques such as padding or timing delays can help protect against such attacks; see Section 8.2.2.

If separate entities provide the oblivious request resource and oblivious target resource, these entities might need an arrangement similar to that between server and proxy for managing denial of service; see Section 8.2.1. It is also necessary to provide confidentiality protection for the unprotected requests and responses, plus protections for traffic analysis; see Section 8.2.2.

An oblivious request resource needs to have a plan for replacing keys. This might include regular replacement of keys, which can be assigned new key identifiers. If an oblivious request resource receives a request that contains a key identifier that it does not understand or that corresponds to a key that has been replaced, the server can respond with an HTTP 422 (Unprocessable Content) status code.

A server can also use a 422 status code if the server has a key that corresponds to the key identifier, but the encapsulated request cannot be successfully decrypted using the key.

A server MUST ensure that the HPKE keys it uses are not valid for any other protocol that uses HPKE with the "message/bhttp request" label. Designers of protocols that reuse this encapsulation format, especially new versions of this protocol, can ensure key diversity by choosing a different label in their use of HPKE. The "message/bhttp response" label was chosen for symmetry only as it provides key diversity only within the HPKE context created using the "message/bhttp request" label; see Section 11.

A server is responsible for either rejecting replayed requests or ensuring that the effect of replays does not adversely affect clients or resources; see Section 8.4.

8.4. Replay Attacks

Encapsulated requests can be copied and replayed by the oblivious proxy resource. The design of oblivious HTTP does not assume that the oblivious proxy resource will not replay requests. In addition, if a client sends an encapsulated request in TLS early data (see Section 8 of [TLS] and [RFC8470]), a network-based adversary might be able to cause the request to be replayed. In both cases, the effect of a replay attack and the mitigations that might be employed are similar to TLS early data.

A client or oblivious proxy resource MUST NOT automatically attempt to retry a failed request unless it receives a positive signal indicating that the request was not processed or forwarded. The HTTP/2 REFUSED_STREAM error code (Section 8.1.4 of [RFC7540]), the HTTP/3 H3_REQUEST_REJECTED error code (Section 8.1 of [QUIC-HTTP]), or a GOAWAY frame with a low enough identifier (in either protocol version) are all sufficient signals that no processing occurred. Connection failures or interruptions are not sufficient signals that no processing occurred.

The anti-replay mechanisms described in Section 8 of [TLS] are generally applicable to oblivious HTTP requests. The encapsulated keying material (or enc) can be used in place of a nonce to uniquely identify a request. This value is a high-entropy value that is freshly generated for every request, so two valid requests will have different values with overwhelming probability.

The mechanism used in TLS for managing differences in client and server clocks cannot be used as it depends on being able to observe previous interactions. Oblivious HTTP explicitly prevents such linkability.

The considerations in [RFC8470] as they relate to managing the risk of replay also apply, though there is no option to delay the processing of a request.

Limiting requests to those with safe methods might not be satisfactory for some applications, particularly those that involve the submission of data to a server. The use of idempotent methods might be of some use in managing replay risk, though it is important to recognize that different idempotent requests can be combined to be not idempotent.

Even without replay prevention, the server-chosen response_nonce field ensures that responses have unique AEAD keys and nonces even when requests are replayed.

8.4.1. Use of Date for Anti-Replay

Clients SHOULD include a Date header field in encapsulated requests. Though HTTP requests often do not include a Date header field, the value of this field might be used by a server to limit the amount of requests it needs to track when preventing replays.

A server can maintain state for requests for a small window of time over which it wishes to accept requests. The server then rejects requests if the request is the same as one that was previously answered within that time window. Servers can reject requests outside of this window and signal that clients might retry with a different Date header field; see Section 4 of [REQUEST-DATE]. Servers can identify duplicate requests using the encapsulation (enc) value.

Servers SHOULD allow for the time it takes requests to arrive from the client, with a time window that is large enough to allow for differences in the clock of clients and servers. How large a time window is needed could depend on the population of clients that the server needs to serve.

Servers MUST NOT treat the time window as secret information. An attacker can actively probe the server with specially crafted request timestamps to determine the time window over which the server will accept responses.

[REQUEST-DATE] contains further considerations for the use of the Date request header field. This includes the way in which clients might correct for clock skew and the privacy considerations arising from that usage. Servers that reject requests on the basis of the Date request header field SHOULD implement the feedback mechanism in Section 4 of [REQUEST-DATE] to support clock correction by clients.

8.5. Post-Compromise Security

This design does not provide post-compromise security for responses.

A client only needs to retain keying material that might be used compromise the confidentiality and integrity of a response until that response is consumed, so there is negligible risk associated with a client compromise.

A server retains a secret key that might be used to remove protection from messages over much longer periods. A server compromise that provided access to the oblivious request resource secret key could allow an attacker to recover the plaintext of all requests sent toward affected keys and all of the responses that were generated.

Even if server keys are compromised, an adversary cannot access messages exchanged by the client with the oblivious proxy resource as messages are protected by TLS. Use of a compromised key also requires that the oblivious proxy resource cooperate with the attacker or that the attacker is able to compromise these TLS connections.

The total number of affected messages affected by server key compromise can be limited by regular rotation of server keys.

9. Privacy Considerations

One goal of this design is that independent client requests are only linkable by the chosen key configuration. The oblivious proxy and request resources can link requests using the same key configuration by matching KeyConfig.key_id, or, if the oblivious target resource is willing to use trial decryption, a limited set of key configurations that share an identifier. An oblivious proxy can link requests using the public key corresponding to KeyConfig.key_id.

Request resources are capable of linking requests depending on how KeyConfigs are produced by servers and discovered by clients. Specifically, servers can maliciously construct key configurations to track individual clients. A specific method for a client to acquire key configurations is not included in this specification. Clients need to consider these tracking vectors when choosing a discovery method. Applications using this design should provide accommodations to mitigate tracking using key configurations. [CONSISTENCY] provides an analysis of the options for ensuring the key configurations are consistent between different clients.

10. Operational and Deployment Considerations

This section discusses various operational and deployment considerations.

10.1. Performance Overhead

Using Oblivious HTTP adds both cryptographic and latency to requests relative to a simple HTTP request-response exchange. Deploying proxy services that are on path between clients and servers avoids adding significant additional delay due to network topology. A study of a similar system [ODoH] found that deploying proxies close to servers was most effective in minimizing additional latency.

10.2. Resource Mappings

This protocol assumes a fixed, one-to-one mapping between the Oblivious Proxy Resource and the Oblivious Request Resource. This means that any encapsulated request sent to the Oblivious Proxy Resource will always be forwarded to the Oblivious Request Resource. This constraint was imposed to simplify proxy configuration and mitigate against the Oblivious Proxy Resource being used as a generic proxy for unknown Oblivious Request Resources. The proxy will only forward for Oblivious Request Resources that it has explicitly configured and allowed.

It is possible for a server to be configured with multiple Oblivious Proxy Resources, each for a different Oblivious Request Resource as needed.

10.3. Network Management

Oblivious HTTP might be incompatible with network interception regimes, such as those that rely on configuring clients with trust anchors and intercepting TLS connections. While TLS might be intercepted successfully, interception middleboxes devices might not receive updates that would allow Oblivious HTTP to be correctly identified using the media types defined in Section 7.

Oblivious HTTP has a simple key management design that is not trivially altered to enable interception by intermediaries. Clients that are configured to enable interception might choose to disable Oblivious HTTP in order to ensure that content is accessible to middleboxes.

11. Repurposing the Encapsulation Format

The encapsulated payload of an OHTTP request and response is a binary HTTP message [BINARY]. Client and target agree on this encapsulated payload type by specifying the media type "message/bhttp" in the HPKE encapsulation info string and HPKE export context string for request and response encapsulation, respectively.

Future specifications may repurpose the encapsulation mechanism described in Section 5, provided that the content type of the encapsulated payload is appropriately reflected in the HPKE info and context strings. For example, if a future specification were to use the encapsulation mechanism in this specification for DNS messages, identified by the "application/dns-message" media type, then the HPKE info string SHOULD be "application/dns-message request" for request encapsulation, and the HPKE export context string should be "application/dns-message response" for response encapsulation.

12. IANA Considerations

Please update the "Media Types" registry at https://www.iana.org/assignments/media-types with the registration information in Section 7 for the media types "message/ohttp-req", "message/ohttp-res", and "application/ohttp-keys".

13. References

13.1. Normative References

[BINARY]
Thomson, M. and C. A. Wood, "Binary Representation of HTTP Messages", Work in Progress, Internet-Draft, draft-ietf-httpbis-binary-message-01, , <https://datatracker.ietf.org/doc/html/draft-ietf-httpbis-binary-message-01>.
[HPKE]
Barnes, R. L., Bhargavan, K., Lipp, B., and C. A. Wood, "Hybrid Public Key Encryption", Work in Progress, Internet-Draft, draft-irtf-cfrg-hpke-12, , <https://datatracker.ietf.org/doc/html/draft-irtf-cfrg-hpke-12>.
[HTTP]
Fielding, R. T., Nottingham, M., and J. Reschke, "HTTP Semantics", Work in Progress, Internet-Draft, draft-ietf-httpbis-semantics-19, , <https://datatracker.ietf.org/doc/html/draft-ietf-httpbis-semantics-19>.
[QUIC]
Iyengar, J., Ed. and M. Thomson, Ed., "QUIC: A UDP-Based Multiplexed and Secure Transport", RFC 9000, DOI 10.17487/RFC9000, , <https://www.rfc-editor.org/rfc/rfc9000>.
[QUIC-HTTP]
Bishop, M., "Hypertext Transfer Protocol Version 3 (HTTP/3)", Work in Progress, Internet-Draft, draft-ietf-quic-http-34, , <https://datatracker.ietf.org/doc/html/draft-ietf-quic-http-34>.
[REQUEST-DATE]
Thomson, M., "Using The Date Header Field In HTTP Requests", Work in Progress, Internet-Draft, draft-thomson-httpapi-date-requests-00, , <https://datatracker.ietf.org/doc/html/draft-thomson-httpapi-date-requests-00>.
[RFC2119]
Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, , <https://www.rfc-editor.org/rfc/rfc2119>.
[RFC7540]
Belshe, M., Peon, R., and M. Thomson, Ed., "Hypertext Transfer Protocol Version 2 (HTTP/2)", RFC 7540, DOI 10.17487/RFC7540, , <https://www.rfc-editor.org/rfc/rfc7540>.
[RFC8174]
Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, , <https://www.rfc-editor.org/rfc/rfc8174>.
[RFC8470]
Thomson, M., Nottingham, M., and W. Tarreau, "Using Early Data in HTTP", RFC 8470, DOI 10.17487/RFC8470, , <https://www.rfc-editor.org/rfc/rfc8470>.
[TLS]
Rescorla, E., "The Transport Layer Security (TLS) Protocol Version 1.3", RFC 8446, DOI 10.17487/RFC8446, , <https://www.rfc-editor.org/rfc/rfc8446>.

13.2. Informative References

[ALT-SVC]
Nottingham, M., McManus, P., and J. Reschke, "HTTP Alternative Services", RFC 7838, DOI 10.17487/RFC7838, , <https://www.rfc-editor.org/rfc/rfc7838>.
[CONSISTENCY]
Davidson, A., Finkel, M., Thomson, M., and C. A. Wood, "Key Consistency and Discovery", Work in Progress, Internet-Draft, draft-wood-key-consistency-01, , <https://datatracker.ietf.org/doc/html/draft-wood-key-consistency-01>.
[COOKIES]
Barth, A., "HTTP State Management Mechanism", RFC 6265, DOI 10.17487/RFC6265, , <https://www.rfc-editor.org/rfc/rfc6265>.
[Dingledine2004]
Dingledine, R., Mathewson, N., and P. Syverson, "Tor: The Second-Generation Onion Router", , <https://svn.torproject.org/svn/projects/design-paper/tor-design.html>.
[FORWARDED]
Petersson, A. and M. Nilsson, "Forwarded HTTP Extension", RFC 7239, DOI 10.17487/RFC7239, , <https://www.rfc-editor.org/rfc/rfc7239>.
[ODoH]
Singanamalla, S., Chunhapanya, S., Vavrusa, M., Verma, T., Wu, P., Fayed, M., Heimerl, K., Sullivan, N., and C. A. Wood, "Oblivious DNS over HTTPS (ODoH): A Practical Privacy Enhancement to DNS", , <https://www.petsymposium.org/2021/files/papers/issue4/popets-2021-0085.pdf>.
[ODOH]
Kinnear, E., McManus, P., Pauly, T., Verma, T., and C. A. Wood, "Oblivious DNS Over HTTPS", Work in Progress, Internet-Draft, draft-pauly-dprive-oblivious-doh-10, , <https://datatracker.ietf.org/doc/html/draft-pauly-dprive-oblivious-doh-10>.
[OHTTP-ANALYSIS]
Hoyland, J., "Tamarin Model of Oblivious HTTP", , <https://github.com/cloudflare/ohttp-analysis>.
[PRIO]
Corrigan-Gibbs, H. and D. Boneh, "Prio: Private, Robust, and Scalable Computation of Aggregate Statistics", , <https://crypto.stanford.edu/prio/paper.pdf>.
[RANDOM]
Eastlake 3rd, D., Schiller, J., and S. Crocker, "Randomness Requirements for Security", BCP 106, RFC 4086, DOI 10.17487/RFC4086, , <https://www.rfc-editor.org/rfc/rfc4086>.
[RFC6265]
Barth, A., "HTTP State Management Mechanism", RFC 6265, DOI 10.17487/RFC6265, , <https://www.rfc-editor.org/rfc/rfc6265>.
[RFC7838]
Nottingham, M., McManus, P., and J. Reschke, "HTTP Alternative Services", RFC 7838, DOI 10.17487/RFC7838, , <https://www.rfc-editor.org/rfc/rfc7838>.
[X25519]
Langley, A., Hamburg, M., and S. Turner, "Elliptic Curves for Security", RFC 7748, DOI 10.17487/RFC7748, , <https://www.rfc-editor.org/rfc/rfc7748>.

Appendix A. Complete Example of a Request and Response

A single request and response exchange is shown here. Binary values (key configuration, secret keys, the content of messages, and intermediate values) are shown in hexadecimal. The request and response here are absolutely minimal; the purpose of this example is to show the cryptographic operations.

The oblivious request resource generates a key pair. In this example the server chooses DHKEM(X25519, HKDF-SHA256) and generates an X25519 key pair [X25519]. The X25519 secret key is:

b8f3cea0da634e6b8271f5b8f931d266decdd04c8e09b80cb9878ea90086ed4a

The oblivious request resource constructs a key configuration that includes the corresponding public key as follows:

01002076eae6d5a6c1549a3343d31c0b9b9582470c72ca11607d47f005f8c16b
e3304a00080001000100010003

This key configuration is somehow obtained by the client. Then when a client wishes to send an HTTP request of a GET request to https://example.com, it constructs the following binary HTTP message:

00034745540568747470730b6578616d706c652e636f6d012f

The client then reads the oblivious request resource key configuration and selects a mutually supported KDF and AEAD. In this example, the client selects HKDF-SHA256 and AES-128-GCM. The client then generates an HPKE context that uses the server public key. This context is constructed from the following ephemeral public key:

25845c6ed6802abfd09628b5c677842b10dd53a3aad5775aa3c20cbae1c0cb65

The corresponding private key is:

88894d1fb4e76e215d9d9c87c44d9e0a6053c1c84c836a4106ea547344504658

Applying the Seal operation from the HPKE context produces an encrypted message, allowing the client to construct the following encapsulated request:

010020000100012485c2eee03135aac82e572d51639b2e141102c431d544e346
0a0784a908b41b676943ae0b3ed8eff581ee8be5f47303a314de092e2e27e3f0
2a03b9357de1d414cadacabaa1621cf9

The client then sends this to the oblivious proxy resource in a POST request, which might look like the following HTTP/1.1 request:

POST /request.example.net/proxy HTTP/1.1
Host: proxy.example.org
Content-Type: message/ohttp-req
Content-Length: 78

<content is the encapsulated request above>

The oblivious proxy resource receives this request and forwards it to the oblivious request resource, which might look like:

POST /oblivious/request HTTP/1.1
Host: example.com
Content-Type: message/ohttp-req
Content-Length: 78

<content is the encapsulated request above>

The oblivous request resource receives this request, selects the key it generated previously using the key identifier from the message, and decrypts the message. As this request is directed to the same server, the oblivious request resource does not need to initiate an HTTP request to the oblivious target resource. The request can be served directly by the oblivious target resource, which generates a minimal response (consisting of just a 200 status code) as follows:

0140c8

The response is constructed by extracting a secret from the HPKE context:

50030a0eacaa9c020e60390c573c4f80

The key derivation for the encapsulated response uses both the encapsulated KEM key from the request and a randomly selected nonce. This produces a salt of:

2485c2eee03135aac82e572d51639b2e141102c431d544e3460a0784a908b41b
e29f9834fd61ffa27f494dfea94d9ed5

The salt and secret are both passed to the Extract function of the selected KDF (HKDF-SHA256) to produce a pseudorandom key of:

4fd8939221446411c785dc9dc51a196df43646a7791919248d0c7624c9410e5b

The pseudorandom key is used with the Expand function of the KDF and an info field of "key" to produce a 16-byte key for the selected AEAD (AES-128-GCM):

a0cd40e2e68cd500bfd14275b290f337

With the same KDF and pseudorandom key, an info field of "nonce" is used to generate a 12-byte nonce:

86883bbe97a380ec2fa656f7

The AEAD Seal function is then used to encrypt the response, which is added to the randomized nonce value to produce the encapsulated response:

e29f9834fd61ffa27f494dfea94d9ed543d89abe34977e7d6d5e1d8051e7b3ba
4ff234

The oblivious request resource then constructs a response:

HTTP/1.1 200 OK
Date: Wed, 27 Jan 2021 04:45:07 GMT
Cache-Control: private, no-store
Content-Type: message/ohttp-res
Content-Length: 38

<content is the encapsulated response>

The same response might then be generated by the oblivious proxy resource which might change as little as the Date header. The client is then able to use the HPKE context it created and the nonce from the encapsulated response to construct the AEAD key and nonce and decrypt the response.

Acknowledgments

This design is based on a design for oblivious DoH, described in [ODOH]. David Benjamin and Eric Rescorla made technical contributions.

Authors' Addresses

Martin Thomson
Mozilla
Christopher A. Wood
Cloudflare