Number
123
Author
zzz, str4d, orignal
Created
Thread
http://zzz.i2p/topics/2051
Last updated
Status
Open
Supercedes
110, 120, 121, 122

Overview

This is an update and aggregation of the following 4 proposals:

  • 110 LS2
  • 120 Meta LS2 for massive multihoming
  • 121 Encrypted LS2
  • 122 Unauthenticated service lookup (anycasting)

These proposals are mostly independent, but for sanity we define and use a common format for several of them.

The following proposals are somewhat related:

  • 140 Invisible Multihoming (incompatible with this proposal)
  • 142 New Crypto Template (for new symmetric crypto)
  • ECIES http://zzz.i2p/topics/2418

Proposal

This proposal defines 5 new DatabaseEntry types and the process for storing them to and retrieving them from the network database, as well as the method for signing them and verifying those signatures.

Goals

  • Backwards compatible
  • LS2 Usable with old-style mulithoming
  • No new crypto or primitives required for support
  • Maintain decoupling of crypto and signing; support all current and future versions
  • Enable optional offline signing keys
  • Reduce accuracy of timestamps to reduce fingerprinting
  • Enable new crypto for destinations
  • Enable massive multihoming
  • Fix multiple issues with existing encrypted LS
  • Optional blinding to reduce visibility by floodfills
  • Encrypted supports both single-key and multiple revocable keys
  • Service lookup for easier lookup of outproxies, application DHT bootstrap, and other uses
  • Don't break anything that relies on 32-byte binary destination hashes, e.g. bittorrent
  • Add flexibility to leasesets via properties, like we have in routerinfos.
  • Put published timestamp and variable expiration in header, so it works even if contents are encrypted (don't derive timestamp from earliest lease)
  • All new types live in the same DHT space and same locations as existing leasesets, so that users may migrate from the old LS to LS2, or change among LS2, Meta, and Encrypted, without changing the Destination or hash.
  • An existing Destination may be converted to use offline keys, or back to online keys, without changing the Destination or hash.

Non-Goals / Out-of-scope

  • New DHT rotation algorithm or shared random generation
  • The specific new encryption type and end-to-end encryption scheme to use that new type would be in a separate proposal. No new crypto is specified or discussed here.
  • New encryption for RIs or tunnel building. That would be in a separate proposal.
  • Methods of encryption, transmission, and reception of I2NP DLM / DSM / DSRM messages. Not changing.
  • How to generate and support Meta, including backend inter-router communication, management, failover, and coordination. Support may be added to I2CP, or i2pcontrol, or a new protocol. This may or may not be standardized.
  • How to actually implement and manage longer-expiring tunnels, or cancel existing tunnels. That's extremely difficult, and without it, you can't have a reasonable graceful shutdown.
  • Threat model changes
  • Offline storage format, or methods to store/retrieve/share the data.
  • Implementation details are not discussed here and are left to each project.

Justification

LS2 adds fields for changing encryption type and for future protocol changes.

Encrypted LS2 fixes several security issues with the existing encrypted LS by using asymmetric encryption of the entire set of leases.

Meta LS2 provides flexible, efficient, effective, and large-scale multihoming.

Service Record and Service List provide anycast services such as naming lookup and DHT bootstrapping.

NetDB Data Types

The type numbers are used in the I2NP Database Lookup/Store Messages.

The end-to-end column means is it sent to a Destination in a Garlic Message.

Existing types:

NetDB Data Lookup Type Store Type
any 0 any
LS 1 1
RI 2 0
exploratory 3 DSRM

New types:

NetDB Data Lookup Type Store Type Std. LS2 Header? Sent end-to-end?
LS2 1 3 yes yes
Encrypted LS2 1 5 no no
Meta LS2 1 7 yes no
Service Record n/a 9 yes no
Service List 4 11 no no

Notes

  • Lookup types are currently bits 3-2 in the Database Lookup Message. Any additional types would require use of bit 4.
  • All store types are odd since upper bits in the Database Store Message type field are ignored by old routers. We would rather have the parse fail as an LS than as a compressed RI.
  • Should be type be explicit or implicit or neither in the data covered by the signature?

Lookup/Store process

Types 3, 5, and 7 may be returned in response to a standard leaseset lookup (type 1). Type 9 is never returned in response to a lookup. Types 11 is returned in response to a new service lookup type (type 11).

Only type 3 may be sent in a client-to-client Garlic message.

Format

Types 3, 7, and 9 all have a common format:

Standard LS2 Header
- as defined below

Type-Specific Part
- as defined below in each part

Standard LS2 Signature:
- Length as implied by sig type of signing key

Type 5 (Encrypted) does not start with a Destination and has a different format. See below.

Type 11 (Service List) is an aggregation of several Service Records and has a different format. See below.

Privacy/Security Considerations

TBD

Standard LS2 Header

Types 3, 7, and 9 use the standard LS2 header, specified below:

Format

Standard LS2 Header:
- Type (1 byte)
  Not actually in header, but part of data covered by signature.
  Take from field in Database Store Message.
- Destination (387+ bytes)
- Published timestamp (4 bytes, seconds since epoch, rolls over in 2106)
- Expires (2 bytes) (offset from published timestamp in seconds, 18.2 hours max)
- Flags (2 bytes)
  Bit order: 15 14 ... 3 2 1 0
  Bit 0: If 0, no offline keys; if 1, offline keys
  Bit 1: If 0, a standard published leaseset.
         If 1, an unpublished leaseset. Should not be flooded, published, or
         sent in response to a query. If this leaseset expires, do not query the
         netdb for a new one.
  Bits 2-15: set to 0 for compatibility with future uses
- If flag indicates offline keys, the offline signature section:
  Expires timestamp (4 bytes, seconds since epoch, rolls over in 2106)
  Transient sig type (2 bytes)
  Transient signing public key (length as implied by sig type)
  Signature of expires timestamp, transient sig type, and public key, by the destination public key,
  length as implied by destination public key sig type.
  This section can, and should, be generated offline.

Justification

  • Unpublished/published: For use when sending a database store end-to-end, the sending router may wish to indicate that this leaseset should not be sent to others. We currently use heuristics to maintain this state.
  • Published: Replaces the complex logic required to determine the 'version' of the leaseset. Currently, the version is the expiration of the last-expiring lease, and a publishing router must increment that expiration by at least 1ms when publishing a leaseset that only removes an older lease.
  • Expires: Allows for an expiration of a netdb entry to be earlier than that of its last-expiring leaseset. May not be useful for LS2, where leasesets are expected to remain with a 11-minute maximum expiration, but for other new types, it is necessary (see Meta LS and Service Record below).
  • Offline keys are optional, to reduce initial/required implementation complexity.

Issues

  • Could reduce timestamp accuracy even more (10 minutes?) but would have to add version number. This could break multihoming, unless we have order preserving encryption? Probably can't do without timestamps at all.
  • Alternative: 3 byte timestamp (epoch / 10 minutes), 1-byte version, 2-byte expires
  • Is type explicit or implicit in data / signature? "Domain" constants for signature?

Notes

  • Routers should not publish a LS more than once a second. If they do, they must artificially increment the published timestamp by 1 over the previously published LS.
  • Router implementations could cache the transient keys and signature to avoid verification every time. In particular, floodfills, and routers at both ends of long-lived connections, could benefit from this.
  • Offline keys and signature are only appropriate for long-lived destinations, i.e. servers, not clients.

New DatabaseEntry types

LeaseSet 2

Changes from existing LeaseSet:

  • Add published timestamp, expires timestamp, flags, and properties
  • Add encryption type
  • Remove revocation key
Lookup with:
Standard LS flag (1)
Store with:
Standard LS2 type (3)
Store at:
Hash of destination, with daily rotation, as for LS 1
Typical expiration:
10 minutes, as in a regular LS.
Published by:
Destination

Format

Standard LS2 Header as specified above

Properties:
- A Mapping, for future use, no current plans.

Standard LS2 Type-Specific Part
- Properties (Mapping as specified in common structures spec, 2 zero bytes if none)
- Number of key sections to follow (1 byte, max TBD)
- Key sections:
  - Encryption type (2 bytes)
  - Encryption key length (2 bytes)
    This is explicit, so floodfills can parse LS2 with unknown encryption types.
  - Encryption key (number of bytes specified)
- Number of lease2s (1 byte)
- Lease2s (40 bytes each)
  These are leases, but with a 4-byte instead of an 8-byte expiration,
  seconds since the epoch (rolls over in 2106)

Standard LS2 Signature:
- Signature
  If flag indicates offline keys, this is signed by the transient pubkey, otherwise, by the destination pubkey
  Length as implied by sig type of signing key
  The signature is of everything above.

Justification

  • Properties: Future expansion and flexibility. Placed first in case necessary for parsing of the remaining data.
  • Multiple encryption type/public key pairs are to ease transition to new encryption types. The other way to do it is to publish multiple leasesets, possibly using the same tunnels, as we do now for DSA and EdDSA destinations. Identification of the incoming encryption type on a tunnel may be done with the existing session tag mechanism, and/or trial decryption using each key. Lengths of the incoming messages may also provide a clue.

Discussion

This proposal continues to use the public key in the leaseset for the end-to-end encryption key, and leaves the public key field in the Destination unused, as it is now. The encryption type is not specified in the Destination key certificate, it will remain 0.

A rejected alternative is to specify the encryption type in the Destination key certificate, use the public key in the Destination, and not use the public key in the leaseset. We do not plan to do this.

Benefits of LS2:

  • Location of actual public key doesn't change.
  • Encryption type, or public key, may change without changing the Destination.
  • Removes unused revocation field
  • Basic compatibility with other DatabaseEntry types in this proposal
  • Allow multiple encryption types

Drawbacks of LS2:

  • Location of public key and encryption type differs from RouterInfo
  • Maintains unused public key in leaseset
  • Requires implementation across the network; in the alternative, experimental encryption types may be used, if allowed by floodfills (but see related proposals 136 and 137 about support for experimental sig types). The alternative proposal could be easier to implement and test for experimental encryption types.

New Encryption Issues

Some of this is out-of-scope for this proposal, but putting notes here for now as we don't have a separate encryption proposal yet. See also the ECIES thread on zzz.i2p.

  • The encryption type represents the combination of curve, key length, and end-to-end scheme, including KDF and MAC, if any.
  • We have included a key length field, so that the LS2 is parsable and verifiable by the floodfill even for unknown encryption types.
  • The first new encryption type to be proposed will probably be ECIES/X25519. How it's used end-to-end (either a slightly modified version of ElGamal/AES+SessionTag or something completely new, e.g. ChaCha/Poly) will be specified in one or more separate proposals. See also the ECIES thread on zzz.i2p.

Notes

  • 8-byte expiration in leases changed to 4 bytes. Alternatives: 2-byte offset from the published timestamp in seconds? Or 4-byte offset in milliseconds?
  • If we ever implement revocation, we can do it with an expires field of zero, or zero leases, or both. No need for a separate revocation key.

Encrypted LS2

Goals:

  • Add blinding
  • Allow multiple sig types
  • Don't require any new crypto primitives
  • Optionally encrypt to each recipient, revokable
  • Support encryption of Standard LS2 and Meta LS2 only

Encrypted LS2 is never sent in an end-to-end garlic message. Use the standard LS2 as above.

You can't use a b32 for an encrypted LS2, as you don't have the non-blinded public key. We need a new "b33" format, or use one of the four unused bits at the end of b32 to indicate it's blinded. You can't use an encrypted LS2 for bittorrent, because of compact announce replies.

Changes from existing encrypted LeaseSet:

  • Encrypt the whole thing for security
  • Securely encrypt, not with AES only.
  • Encrypt to each recipient
Lookup with:
Standard LS flag (1)
Store with:
Encrypted LS2 type (5)
Store at:
Hash of blinded sig type and public key, with daily rotation
Typical expiration:
10 minutes, as in a regular LS.
Published by:
Destination

Definitions

We define the following functions corresponding to the cryptographic building blocks used for encrypted LS2:

CSRNG(n)

n-byte output from a cryptographically-secure random number generator.

In addition to the requirement of CSRNG being cryptographically-secure (and thus suitable for generating key material), it MUST be instantiated such that it is safe for some n-byte output to be used for key material when the byte sequences immediately preceding and following it are exposed on the network (such as in a salt, or encrypted padding). Implementations that rely on a potentially-untrustworthy source should hash any output that is to be exposed on the network [PRNG-REFS].

H(p, d)

A cryptographic hash function that takes a personalisation string p and data d, and produces an output of length HASH_LEN bytes. The hash function should be preimage- and collision-resistant.

Instantiated with SHA-256 (implying HASH_LEN = 32) as follows:

H(p, d) := SHA-256(p || d)
STREAM

The ChaCha20 stream cipher as specified in [RFC-7539-S2.4], with the initial counter set to 1. S_KEY_LEN = 32 and S_IV_LEN = 12.

ENCRYPT(k, iv, plaintext)

Encrypts plaintext using the cipher key k, and nonce iv which MUST be unique for the key k. Returns a ciphertext that is the same size as the plaintext.

The entire ciphertext must be indistinguishable from random if the key is secret.

DECRYPT(k, iv, ciphertext)
Decrypts ciphertext using the cipher key k, and nonce iv. Returns the plaintext.
SIG

The Ed25519 signature scheme (corresponding to SigType 7) with key-blinding. It has the following functions:

DERIVE_PUBLIC(privkey)
Returns the public key corresponding to the given private key.
SIGN(privkey, m)
Returns a signature by the private key privkey over the given message m.
VERIFY(pubkey, m, sig)
Verifies the signature sig against the public key pubkey and message m. Returns true if the signature is valid, false otherwise.

It must also support the following key blinding operations:

BLIND_PRIVKEY(privkey, blind)
Blinds a private key. Blinding is roughly as specified in Tor's rend-spec-v3 appendices A.1 and A.2. TODO
BLIND_PUBKEY(pubkey, blind)

Blinds a public key, such that for a given keypair (privkey, pubkey) the following relationship holds:

BLIND_PUBKEY(pubkey, blind) == DERIVE_PUBLIC(BLIND_PRIVKEY(privkey, blind))

Blinding is roughly as specified in Tor's rend-spec-v3 appendices A.1 and A.2. TODO

DH

X25519 public key agreement system. Private keys of 32 bytes, public keys of 32 bytes, produces outputs of 32 bytes. DH_PUBKEY_LEN = 32. It has the following functions:

GENERATE_PRIVATE()
Generates a new private key.
DERIVE_PUBLIC(privkey)
Returns the public key corresponding to the given private key.
AGREE(privkey, pubkey)
Generates a shared secret from the given private and public keys.
KDF(salt, ikm, info, n)

A cryptographic key derivation function which takes some input key material ikm (which should have good entropy but is not required to be a uniformly random string), a salt of length SALT_LEN bytes, and a context-specific 'info' value, and produces an output of n bytes suitable for use as key material.

Instantiated with HKDF as specified in [RFC-5869], using the hash function SHA-256. This means that SALT_LEN can be at most 32.

Format

The encrypted LS2 format consists of three nested layers:

  • An outer layer containing the necessary plaintext information for storage and retrieval.
  • A middle layer that handles client authentication.
  • An inner layer that contains the actual LS2 data.

The overall format looks like:

Layer 0 data + Enc(layer 1 data + Enc(layer 2 data)) + Signature

Note that encrypted LS2 is blinded. The Destination is not in the header. DHT storage location is SHA-256(sig type || blinded public key), and rotated daily.

Does NOT use the standard LS2 header specified above.

Layer 0 (outer)

Type

1 byte

Not actually in header, but part of data covered by signature. Take from field in Database Store Message.

Blinded Public Key Sig Type
2 bytes
Blinded Public Key
Length as implied by sig type
Signature

Length as implied by signing key sig type

Of destination by blinded public key?

Published timestamp

4 bytes

Seconds since epoch, rolls over in 2106

Expires

2 bytes

Offset from published timestamp in seconds, 18.2 hours max

Flags

2 bytes

Bit order: 15 14 ... 3 2 1 0

Bit 0: If 0, no offline keys; if 1, offline keys

Other bits: set to 0 for compatibility with future uses

Transient key data

Present if flag indicates offline keys

Expires timestamp

4 bytes

Seconds since epoch, rolls over in 2106

Transient sig type
2 bytes
Transient signing public key
Length as implied by sig type
Signature

Length as implied by blinded public key sig type

Over expires timestamp, transient sig type, and transient public key.

Verified with the blinded public key.

lenOuterCiphertext
2 bytes
outerCiphertext

lenOuterCiphertext bytes

Encrypted layer 1 data. See below for key derivation and encryption algorithms.

Signature

Length as implied by sig type of the signing key used

The signature is of everything above.

If the flag indicates offline keys, the signature is verified with the transient public key. Otherwise, the signature is verified with the blinded public key.

Layer 1 (middle)

Flags

1 byte

Bit order: 76543210

Bit 0: 0 for everybody, 1 for per-client, auth section to follow

Bits 3-1: Authentication scheme, 0 for the scheme specified below

Bits 7-4: Unused, set to 0 for future compatibility

DH client auth data

Present if flag bit 0 is set to 1 and flag bits 3-1 are set to 0.

ephemeralPublicKey
DH_PUBKEY_LEN bytes
lenAuthClient

2 bytes

Number of authClient entries to follow

authClient

Authorization data for a single client

clientID_i
8 bytes
clientCookie_i
32 bytes

See below for per-client authorization algorithm.

innerCiphertext

Length implied by lenOuterCiphertext (whatever data remains)

Encrypted layer 2 data. See below for key derivation and encryption algorithms.

Layer 2 (inner)

Type

1 byte

Either 3 (LS2) or 7 (Meta LS2)

Data

LeaseSet2 data for the given type.

Includes the header and signature.

Blinding Key Derivation

We propose the following scheme for key blinding, based on Ed25519.

(This is an ECC group, so remember that scalar multiplication is the trapdoor function, and it's defined in terms of iterated point addition. See the Ed25519 paper [ED25519-REFS] for a fairly clear writeup.)

Copied from Tor rend-spec-v3.txt appendix A.2 which has similar design goals [TOR-REND-SPEC-V3].

Changes for I2P TODO

Let B be the ed25519 basepoint as found in section 5 of [ED25519-B-REF]:
    B = (15112221349535400772501151409588531511454012693041857206046113283949847762202,
         46316835694926478169428394003475163141307993866256225615783033603165251855960)

Assume B has prime order l, so lB=0. Let a master keypair be written as
(a,A), where a is the private key and A is the public key (A=aB).

To derive the key for a nonce N and an optional secret s, compute the
blinding factor like this:

         h = H(BLIND_STRING | A | s | B | N)
         BLIND_STRING = "Derive temporary signing key" | INT_1(0)
         N = "key-blind" | INT_8(period-number) | INT_8(period_length)
         B = "(1511[...]2202, 4631[...]5960)"

then clamp the blinding factor 'h' according to the ed25519 spec:

         h[0] &= 248;
         h[31] &= 63;
         h[31] |= 64;

and do the key derivation as follows:

    private key for the period:

         a' = h a mod l
         RH' = SHA-512(RH_BLIND_STRING | RH)[:32]
         RH_BLIND_STRING = "Derive temporary signing key hash input"

    public key for the period:

         A' = h A = (ha)B

Generating a signature of M: given a deterministic random-looking r
(see EdDSA paper), take R=rB, S=r+hash(R,A',M)ah mod l. Send signature
(R,S) and public key A'.

Verifying the signature: Check whether SB = R+hash(R,A',M)A'.

(If the signature is valid,
     SB = (r + hash(R,A',M)ah)B
        = rB + (hash(R,A',M)ah)B
        = R + hash(R,A',M)A' )

This boils down to regular Ed25519 with key pair (a', A').

See [KEYBLIND-REFS] for an extensive discussion on this scheme and possible alternatives. Also, see [KEYBLIND-PROOF] for a security proof of this scheme.

Encryption and processing

Derivation of subcredentials

As part of the blinding process, we need to ensure that an encrypted LS2 can only be decrypted by someone who knows the corresponding Destination. To achieve this, we derive a credential from the Destination:

credential = H("credential", Destination)

The personalization string ensures that the credential does not collide with any hash used as a DHT lookup key, such as the plain Destination hash.

For a given blinded key, we can then derive a subcredential:

subcredential = H("subcredential", credential || blindedPublicKey)

The subcredential is included in the key derivation processes below, which binds those keys to knowledge of the Destination.

Layer 1 encryption

First, the input to the key derivation process is prepared:

outerInput = subcredential || publishedTimestamp

Next, a random salt is generated:

outerSalt = CSRNG(SALT_LEN)

Then the key used to encrypt layer 1 is derived:

keys = KDF(outerSalt, outerInput, "ELS2_L1K", S_KEY_LEN + S_IV_LEN)
outerKey = keys[0..S_KEY_LEN]
outerIV = keys[S_KEY_LEN..(S_KEY_LEN+S_IV_LEN)]

Finally, the layer 1 plaintext is encrypted and serialized:

outerCiphertext = outerSalt || ENCRYPT(outerKey, outerIV, outerPlaintext)

Layer 1 decryption

The salt is parsed from the layer 1 ciphertext:

outerSalt = outerCiphertext[0..SALT_LEN]

Then the key used to encrypt layer 1 is derived:

outerInput = subcredential || publishedTimestamp
keys = KDF(outerSalt, outerInput, "ELS2_L1K", S_KEY_LEN + S_IV_LEN)
outerKey = keys[0..S_KEY_LEN]
outerIV = keys[S_KEY_LEN..(S_KEY_LEN+S_IV_LEN)]

Finally, the layer 1 ciphertext is decrypted:

outerPlaintext = DECRYPT(outerKey, outerIV, outerCiphertext[SALT_LEN..])

Layer 2 encryption

When client authorization is enabled, authCookie is calculated as described below. When client authorization is disabled, authCookie is the zero-length byte array.

Encryption proceeds in a similar fashion to layer 1:

innerInput = authCookie || subcredential || publishedTimestamp
innerSalt = CSRNG(SALT_LEN)
keys = KDF(innerSalt, innerInput, "ELS2_L2K", S_KEY_LEN + S_IV_LEN)
innerKey = keys[0..S_KEY_LEN]
innerIV = keys[S_KEY_LEN..(S_KEY_LEN+S_IV_LEN)]
innerCiphertext = innerSalt || ENCRYPT(innerKey, innerIV, innerPlaintext)

Layer 2 decryption

When client authorization is enabled, authCookie is calculated as described below. When client authorization is disabled, authCookie is the zero-length byte array.

Decryption proceeds in a similar fashion to layer 1:

innerInput = authCookie || subcredential || publishedTimestamp
innerSalt = innerCiphertext[0..SALT_LEN]
keys = KDF(innerSalt, innerInput, "ELS2_L2K", S_KEY_LEN + S_IV_LEN)
innerKey = keys[0..S_KEY_LEN]
innerIV = keys[S_KEY_LEN..(S_KEY_LEN+S_IV_LEN)]
innerPlaintext = DECRYPT(innerKey, innerIV, innerCiphertext[SALT_LEN..])

Per-client authorization

When client authorization is enabled for a Destination, the server maintains a list of clients they are authorizing to decrypt the encrypted LS2 data. The data stored per-client depends on the authorization mechanism, and includes some form of key material that each client generates and sends to the server via a secure out-of-band mechanism.

There are two current alternatives for implementing per-client authorization:

DH client authorization

Each client generates a DH keypair [csk_i, cpk_i], and sends the public key cpk_i to the server.

Server processing

The server generates a new authCookie and an ephemeral DH keypair:

authCookie = CSRNG(32)
esk = DH.GENERATE_PRIVATE()
epk = DH.DERIVE_PUBLIC(esk)

Then for each authorized client, the server encrypts authCookie to its public key:

sharedSecret = DH.AGREE(esk, cpk_i)
authInput = sharedSecret || cpk_i || subcredential || publishedTimestamp
okm = KDF(epk, authInput, "ELS2_XCA", S_KEY_LEN + S_IV_LEN + 8)
clientKey_i = okm[0..S_KEY_LEN]
clientIV_i = okm[S_KEY_LEN..(S_KEY_LEN+S_IV_LEN)]
clientID_i = okm[(S_KEY_LEN+S_IV_LEN)..(S_KEY_LEN+S_IV_LEN+8)]
clientCookie_i = ENCRYPT(clientKey_i, clientIV_i, authCookie)

The server places each [clientID_i, clientCookie_i] tuple into layer 1 of the encrypted LS2, along with epk.

Client processing

The client uses its private key to derive its expected client identifier clientID_i, encryption key clientKey_i, and encryption IV clientIV_i:

sharedSecret = DH.AGREE(csk_i, epk)
authInput = sharedSecret || cpk_i || subcredential || publishedTimestamp
okm = KDF(epk, authInput, "ELS2_XCA", S_KEY_LEN + S_IV_LEN + 8)
clientKey_i = okm[0..S_KEY_LEN]
clientIV_i = okm[S_KEY_LEN..(S_KEY_LEN+S_IV_LEN)]
clientID_i = okm[(S_KEY_LEN+S_IV_LEN)..(S_KEY_LEN+S_IV_LEN+8)]

Then the client searches the layer 1 authorization data for an entry that contains clientID_i. If a matching entry exists, the client decrypts it to obtain authCookie:

authCookie = DECRYPT(clientKey_i, clientIV_i, clientCookie_i)

Pre-shared key client authorization

Each client generates a secret 32-byte key psk_i, and sends it to the server.

Server processing

The server generates a new authCookie and salt:

authCookie = CSRNG(32)
authSalt = CSRNG(SALT_LEN)

Then for each authorized client, the server encrypts authCookie to its pre-shared key:

authInput = psk_i || subcredential || publishedTimestamp
okm = KDF(authSalt, authInput, "ELS2PSKA", S_KEY_LEN + S_IV_LEN + 8)
clientKey_i = okm[0..S_KEY_LEN]
clientIV_i = okm[S_KEY_LEN..(S_KEY_LEN+S_IV_LEN)]
clientID_i = okm[(S_KEY_LEN+S_IV_LEN)..(S_KEY_LEN+S_IV_LEN+8)]
clientCookie_i = ENCRYPT(clientKey_i, clientIV_i, authCookie)

The server places each [clientID_i, clientCookie_i] tuple into layer 1 of the encrypted LS2, along with authSalt.

Client processing

The client uses its pre-shared key to derive its expected client identifier clientID_i, encryption key clientKey_i, and encryption IV clientIV_i:

authInput = psk_i || subcredential || publishedTimestamp
okm = KDF(authSalt, authInput, "ELS2PSKA", S_KEY_LEN + S_IV_LEN + 8)
clientKey_i = okm[0..S_KEY_LEN]
clientIV_i = okm[S_KEY_LEN..(S_KEY_LEN+S_IV_LEN)]
clientID_i = okm[(S_KEY_LEN+S_IV_LEN)..(S_KEY_LEN+S_IV_LEN+8)]

Then the client searches the layer 1 authorization data for an entry that contains clientID_i. If a matching entry exists, the client decrypts it to obtain authCookie:

authCookie = DECRYPT(clientKey_i, clientIV_i, clientCookie_i)

Security considerations

Both of the client authorization mechanisms above provide privacy for client membership. An entity that only knows the Destination can see how many clients are subscribed at any time, but cannot track which clients are being added or revoked.

Servers SHOULD randomize the order of clients each time they generate an encrypted LS2, to prevent clients learning their position in the list and inferring when other clients have been added or revoked.

A server MAY choose to hide the number of clients that are subscribed by inserting random entries into the list of authorization data.

Advantages of DH client authorization
  • Security of the scheme is not solely dependent on the out-of-band exchange of client key material. The client's private key never needs to leave their device, and so an adversary that is able to intercept the out-of-band exchange, but cannot break the DH algorithm, cannot decrypt the encrypted LS2, or determine how long the client is given access.
Downsides of DH client authorization
  • Requires N + 1 DH operations on the server side for N clients.
  • Requires one DH operation on the client side.
Advantages of PSK client authorization
  • Requires no DH operations.
Downsides of PSK client authorization
  • Security of the scheme is critically dependent on the out-of-band exchange of client key material. An adversary that intercepts the exchange for a particular client can decrypt any subsequent encrypted LS2 for which that client is authorized, as well as determine when the client's access is revoked.

Issues

  • Blinding spec TODO
  • Use AES instead of ChaCha20?
  • If we care about speed, we could use keyed-BLAKE2b instead. It has an output size large enough to accommodate the largest n we require (or we can call it once per desired key with a counter argument). BLAKE2b is much faster than SHA-256, and keyed-BLAKE2b would reduce the total number of hash function calls. [UNSCIENTIFIC-KDF-SPEEDS]

Notes

  • For multiple clients, encrypted format is probably like GPG/OpenPGP does. Asymmetrically encrypt a symmetric key for each recipient. Data is decrypted with that asymmetric key. See e.g. [RFC-4880-S5.1] IF we can find an algorithm that's small and fast.
    • Can we use a shortened version of our current ElGamal, which is 222 bytes in and 514 bytes out? That's a little long for each record.
  • For a single client, we could just ElG encrypt the whole leaseset, 514 bytes isn't so bad.
  • If we want to specify the encryption format in the clear, we could have an identifier just before the encrypted data, or in the flags.
  • A service using encrypted leasesets would publish the encrypted version to the floodfills. However, for efficiency, it would send unencrypted leasesets to clients in the wrapped garlic message, once authenticated (via whitelist, for example).
  • Floodfills may limit the max size to a reasonable value to prevent abuse.
  • After decryption, several checks should be made, including that the inner timestamp and expiration match those at the top level.

Meta LS2

This is used to replace multihoming. Like any leaseset, this is signed by the creator. This is an authenticated list of destination hashes.

The Meta LS2 is the top of, and possibly intermediate nodes of, a tree structure. It contains a number of entries, each pointing to a LS, LS2, or another Meta LS2 to support massive multihoming. A Meta LS2 may contain a mix of LS, LS2, and Meta LS2 entries. The leaves of the tree are always a LS or LS2. The tree is a DAG; loops are prohibited; clients doing lookups must detect and refuse to follow loops.

A Meta LS2 may have a much longer expiration than a standard LS or LS2. The top level may have an expiration several hours after the publication date. Maximum expiration time will be enforced by floodfills and clients, and is TBD.

The use case for Meta LS2 is massive multihoming, but with no more protection for correlation of routers to leasesets (at router restart time) than is provided now with LS or LS2. This is equivalent to the "facebook" use case, which probably doesn't need correlation protection. This use case probably needs offline keys, which are provided in the standard header at each node of the tree.

The back-end protocol for coordination between the leaf routers, intermediate and master Meta LS signers is not specified here. The requirements are extremely simple - just verify that the peer is up, and publish a new LS every few hours. The only complexity is for picking new publishers for the top-level or intermediate-level Meta LSes on failure.

Mix-and-match leasesets where leases from multiple routers are combined, signed, and published in a single leaseset is documented in proposal 140, "invisible multihoming". This proposal is untenable as written, because streaming connections would not be "sticky" to a single router, see http://zzz.i2p/topics/2335 .

The back-end protocol, and interaction with router and client internals, would be quite complex for invisible multihoming.

To avoid overloading the floodfill for the top-level Meta LS, the expiration should be several hours at least. Clients must cache the top-level Meta LS, and persist it across restarts if unexpired.

We need to define some algorithm for clients to traverse the tree, including fallbacks, so that the usage is dispersed. Some function of hash distance, cost, and randomness. If a node has both LS or LS2 and Meta LS, we need to know when it's allowed to use those leasesets, and when to keep traversing the tree.

Lookup with:
Standard LS flag (1)
Store with:
Meta LS2 type (7)
Store at:
Hash of destination, with daily rotation, as for LS 1
Typical expiration:
Hours. Max 18.2 hours (65535 seconds)
Published by:
"master" Destination or coordinator, or intermediate coordinators

Format

Standard LS2 Header as specified above

Meta LS2 Type-Specific Part
- Properties (Mapping as specified in common structures spec, 2 zero bytes if none)
- Number of entries (1 byte) Maximum TBD
- Entries. Each entry contains: (40 bytes)
  - Hash (32 bytes)
  - Flags (3 bytes)
    TBD. Set all to zero for compatibility with future uses.
    TODO: Use a few bits to (optionally) indicate the type of the LS it is referencing.
    All zeros means don't know.
  - Cost (priority) (1 byte)
  - Expires (4 bytes) (4 bytes, seconds since epoch, rolls over in 2106)
- Number of revocations (1 byte) Maximum TBD
- Revocations: Each revocation contains: (32 bytes)
  - Hash (32 bytes)

Standard LS2 Signature:
- Signature (40+ bytes)
  The signature is of everything above.

Flags and properties: for future use

Notes

  • A distributed service using this would have one or more "masters" with the private key of the service destination. They would (out of band) determine the current list of active destinations and would publish the Meta LS2. For redundancy, multiple masters could multihome (i.e. concurrently publish) the Meta LS2.
  • A distributed service could start with a single destination or use old-style multihoming, then transition to a Meta LS2. A standard LS lookup could return any one of a LS, LS2, or Meta LS2.
  • When a service uses a Meta LS2, it has no tunnels (leases).

Service Record

This is an individual record saying that a destination is participating in a service. It is sent from the participant to the floodfill. It is not ever sent individually by a floodfill, but only as a part of a Service List. The Service Record is also used to revoke participation in a service, by setting the expiration to zero.

This is not a LS2 but it uses the standard LS2 header and signature format.

Lookup with:
n/a, see Service List
Store with:
Service Record type (9)
Store at:
Hash of service name, with daily rotation
Typical expiration:
Hours. Max 18.2 hours (65535 seconds)
Published by:
Destination

Format

Standard LS2 Header as specified above

Service Record Type-Specific Part
- Port (2 bytes) (0 if unspecified)
- Hash of service name (32 bytes)

Standard LS2 Signature:
- Signature (40+ bytes)
  The signature is of everything above.

Notes

  • If expires is all zeros, the floodfill should revoke the record and no longer include it in the service list.
  • Storage: The floodfill may strictly throttle storage of these records and limit the number of records stored per hash and their expiration. A whilelist of hashes may also be used.
  • Any other netdb type at the same hash has priority, so a service record can never overwrite a LS/RI, but a LS/RI will overwrite all service records at that hash.

Service List

This is nothing like a LS2 and uses a different format.

The service list is created and signed by the floodfill. It is unauthenticated in that anybody can join a service by publishing a Service Record to a floodfill.

A Service List contains Short Service Records, not full Service Records. These contain signatures but only hashes, not full destinations, so they cannot be verified without the full destination.

The security, if any, and desirability of service lists is TBD. Floodfills could limit publication, and lookups, to a whitelist of services, but that whitelist may vary based on implementation, or operator preference. It may not be possible to achieve consensus on a common, base whitelist across implementations.

If the service name is included in the service record above, then floodfill operators may object; if only the hash is included, there's no verification, and a service record could "get in" ahead of any other netdb type and get stored in the floodfill.

Lookup with:
Service List lookup type (11)
Store with:
Service List type (11)
Store at:
Hash of service name, with daily rotation
Typical expiration:
Hours, not specified in the list itself, up to local policy
Published by:
Nobody, never sent to floodfill, never flooded.

Format

Does NOT use the standard LS2 header specified above.

- Type (1 byte)
  Not actually in header, but part of data covered by signature.
  Take from field in Database Store Message.
- Hash of the service name (implicit, in the Database Store message)
- Hash of the Creator (floodfill) (32 bytes)
- Published timestamp (8 bytes)

- Number of Short Service Records (1 byte)
- List of Short Service Records:
  Each Short Service Record contains (90+ bytes)
  - Dest hash (32 bytes)
  - Published timestamp (8 bytes)
  - Expires (4 bytes) (offset from published in ms)
  - Flags (2 bytes)
  - Port (2 bytes)
  - Sig length (2 bytes)
  - Signature of dest (40+ bytes)

- Number of Revocation Records (1 byte)
- List of Revocation Records:
  Each Revocation Record contains (86+ bytes)
  - Dest hash (32 bytes)
  - Published timestamp (8 bytes)
  - Flags (2 bytes)
  - Port (2 bytes)
  - Sig length (2 bytes)
  - Signature of dest (40+ bytes)

- Signature of floodfill (40+ bytes)
  The signature is of everything above.

To verify signature of the Service List:

  • prepend the hash of the service name
  • remove the hash of the creator
  • Check signature of the modified contents

To verify signature of each Short Service Record:

  • Fetch destination
  • Check signature of (published timestamp + expires + flags + port + Hash of service name)

To verify signature of each Revocation Record:

  • Fetch destination
  • Check signature of (published timestamp + 4 zero bytes + flags + port + Hash of service name)

Notes

  • We use signature length instead of sig type so we can support unknown signature types.
  • There is no expiration of a service list, recipients may make their own decision based on policy or the expiration of the individual records.
  • Service Lists are not flooded, only individual Service Records are. Each floodfill creates, signs, and caches a Service List. The floodfill uses its own policy for cache time and the maximum number of service and revocation records.

Common Structures Spec Changes Required

TODO

Key Certificates

Out of scope for this proposal. Add to ECIES proposal.

Lease2

Add new structure with 4-byte expiration.

New NetDB Types

Incorporate from above.

Encryption Spec Changes Required

Out of scope for this proposal. Add to ECIES proposal.

I2NP Changes Required

TODO Add note: LS2 can only be published to floodfills with a minimum version.

Database Lookup Message

Add the service list lookup type.

Changes

Flags byte: Lookup type field, currently bits 3-2, expands to bits 4-2.
Lookup type 0x04 is defined as the service list lookup.

Add note: Service list loookup may only be sent to floodfills with a minimum version.
Minimum version is 0.9.38.

Database Store Message

Add all the new store types.

Changes

Type byte: Type field, currently bit 0, expands to bits 3-0.
Type 3 is defined as a LS2 store.
Type 5 is defined as a encrypted LS2 store.
Type 7 is defined as a meta LS2 store.
Type 9 is defined as a service record store.
Type 11 is defined as a service list store.
Other types are undefined and invalid.

Add note: All new types may only be published to floodfills with a minimum version.
Minimum version is 0.9.38.

I2CP Changes Required

I2CP Options

New options interpreted router-side, sent in SessionConfig Mapping:

i2cp.leaseSetType=nnn                       The type of leaseset to be sent in the Create Leaseset Message
                                            Value is the same as the netdb store type in the table above.
                                            Interpreted client-side, but also passed to the router in the
                                            SessionConfig, to declare intent and check support.

i2cp.leaseSetOfflineExpiration=nnn          The expiration of the offline signature, 4 bytes,
                                            seconds since the epoch.

i2cp.leaseSetTransientPublicKey=[type:]b64  The base 64 of the transient private key,
                                            prefixed by an optional sig type number or name,
                                            default DSA_SHA1.
                                            Length as inferred from the sig type

i2cp.leaseSetOfflineSignature=b64           The base 64 of the offline signature.
                                            Length as inferred from the destination signing public key type

New options interpreted client-side:

i2cp.leaseSetType=nnn     The type of leaseset to be sent in the Create Leaseset Message
                          Value is the same as the netdb store type in the table above.
                          Interpreted client-side, but also passed to the router in the
                          SessionConfig, to declare intent and check support.

i2cp.leaseSetEncType=nnn  The encryption type to be used.
                          See proposal 144.

Session Config

Note that for offline signatures, the options i2cp.leaseSetOfflineExpiration, i2cp.leaseSetTransientPublicKey, and i2cp.leaseSetOfflineSignature are required, and the signature is by the transient signing private key.

Request Leaseset Message

Router to client. No changes. The leases are sent with 8-byte timestamps, even if the returned leaseset will be a LS2 with 4-byte timestamps. Note that the response may be a Create Leaseset or Create Leaseset2 Message.

Request Variable Leaseset Message

Router to client. No changes. The leases are sent with 8-byte timestamps, even if the returned leaseset will be a LS2 with 4-byte timestamps. Note that the response may be a Create Leaseset or Create Leaseset2 Message.

Create Leaseset2 Message

Client to router. New message, to use in place of Create Leaseset Message.

Justification

  • For the router to parse the store type, the type must be in the message, unless it is passed to the router before hand in the session config. For for common parsing code, it's easier to have it in the message itself.
  • For the router to know the type and length of the private key, it must be after the lease set, unless the parser knows the type before hand in the session config. For for common parsing code, it's easier to know it from the message itself.
  • The signing private key, previously defined for revocation and unused, was before the leaseset so the type and length was unknown. Clients always set it to the DSA length. For proposal 144, the key may be required, and must match the type of the destination signing key (or transient signing key if offline sigs are used). For the router to know the type and length of the private key, it must be after the lease set, unless the parser knows the type before hand in the session config. For for common parsing code, it's easier to know it from the message itself.

Message Type

The message type for the Create Leaseset2 Message is 40.

Format

Session ID
Type byte: Type of lease set to follow
           Type 1 is a LS
           Type 3 is a LS2
           Type 5 is a encrypted LS2
           Type 7 is a meta LS2
LeaseSet: type specified above
Signing Private Key: type as inferred from the lease set signature
                     (by dest signing key or transient key)
                     Not present for Meta LS2
Encryption Private Key: type as inferred from the public key in the lease set
                        Not present for Meta LS2

Notes

  • Minimum router version is 0.9.38.

Issues

  • More changes are needed to support encrypted and meta LS.

Host Lookup Message

Client to router.

A lookup of a hash will force the router to fetch the Lease Set, so extended results may be returned in the Host Reply Message. However, a lookup of a host name will not force the router to fetch the Lease Set (unless the lookup was for a b32.i2p, which is discouraged, the client side normally converts this to a hash lookup).

To force a Lease Set lookup for a host name lookup, we need a new request type.

Changes

Add request type 3: Host name lookup and request Lease Set lookup.

Host Reply Message

Router to client.

A client doesn't know a priori that a given Hash will resolve to a Meta LS.

If a Host Lookup Message for a Hash yields a Meta LS, the router needs to return one or more Destinations and expirations to the client. Either the client must to the recursive resolution, or the router can do it. Not clear how it should work. For either method, we either need a new flavor of the Host Reply Message, or define a new result code that means what follows is a list of Destinations and expirations.

If the router simply returns a single Destination whose Hash doesn't match that of the lookup, it may fail sanity checks on the client side, and the client has no way to get an alternate if that fails, and has no way to know the expiration time.

There may be similar issues in BOB and SAM.

Changes

If the client version is 0.9.38 or higher, and the result code is 0,
the following extended results are included after the Destination.
These are included no matter what the request type.

5.  LeaseSet type (1 byte)
    0: Unknown
    1: LS 1
    3: LS 2
    7: Meta LS
6.  LeaseSet expiration (4 bytes, seconds since the epoch)
    0 if unknown
7.  Number of encryption types supported (1 byte)
    0 if unknown
8.  That number of encryption types, 2 bytes each
9.  Lease set options, a Mapping, or 2 bytes of zeros if unknown.
10. Flags (2 bytes)
    Bit order: 15 14 13...3210
    Bit 0: 1 for offline keys, 0 if not
    Bits 15-1: Unused, set to 0 for compatibility with future uses
11. If offline keys, the transient key sig type (2 bytes)
12. If offline keys, the transient public key (length as implied by sig type)
13. If LeaseSet type is Meta (7), the number of meta entries to follow (1 byte)
14. If LeaseSet type is Meta (7), the Meta Entries. Each entry contains: (40 bytes)
    - Hash (32 bytes)
    - Flags (3 bytes)
      TBD. Set all to zero for compatibility with future uses.
      TODO: Use a few bits to (optionally) indicate the type of the LS it is referencing.
      All zeros means don't know.
    - Cost (priority) (1 byte)
    - Expires (4 bytes) (4 bytes, seconds since epoch, rolls over in 2106)

Changes to support Meta

How to generate and support Meta, including inter-router communication and coordination, is out of scope for this proposal. Support may be added to I2CP, or i2pcontrol, or a new protocol.

Changes to support Offline Keys

Offline signatures cannot be verified in streaming or repliable datagrams. See sections below.

Private Key File Changes Required

The private key file (eepPriv.dat) format is not an official part of our specifications but it is documented in the Java I2P javadocs http://echelon.i2p/javadoc/net/i2p/data/PrivateKeyFile.html and other implementations do support it. This enables portability of private keys to different implementations.

Changes are necessary to store the transient public key and offline signing information.

Changes

If the signing private key is all zeros, the offline information section follows:

- Expires timestamp (4 bytes, seconds since epoch, rolls over in 2106)
- Sig type of transient Signing Public Key (2 bytes)
- Transient Signing Public key (length as specified by transient sig type)
- Signature of above three fields by offline key (length as specified by destination sig type)
- Transient Signing Private key (length as specified by transient sig type)

Streaming Changes Required

Offline signatures cannot currently be verified in streaming. The change below adds the offline signing block to the options. This avoids having to retrieve this information via I2CP.

Changes

Change option:
Bit:          3
Flag:         SIGNATURE_INCLUDED
Option order: Change from 4 to 5

Add new option:
Bit:          11
Flag:         OFFLINE_SIGNATURE
Option order: 4
Option data:  Variable bytes
Function:     Contains the offline signature block from LS2.
              SIGNATURE_INCLUDED must also be set.
              Expires timestamp (4 bytes, seconds since epoch, rolls over in 2106)
              Transient sig type (2 bytes)
              Transient signing public key (length as implied by sig type)
              Signature of expires timestamp, transient sig type, and public key, by the destination public key,
              length as implied by destination public key sig type.

Add information about transient keys to the Variable Length Signature Notes section:
The offline signature option does not needed to be added for a CLOSE packet if
a SYN packet containing the option was previously acked.
More info TODO

Notes

  • Alternative is to just add a flag, and retrieve the transient public key via I2CP (See Host Lookup / Host Reply Message sections above)

Repliable Datagram Changes Required

Offline signatures cannot be verified in the repliable datagram processing. Needs a flag to indicate offline signed but there's no place to put a flag. Will require a completely new protocol number and format.

Changes

Define new protocol 19 - Repliable datagram with options?
- Destination (387+ bytes)
- Flags (2 bytes)
  Bit order: 15 14 ... 3 2 1 0
  Bit 0: If 0, no offline keys; if 1, offline keys
  Bits 1-15: set to 0 for compatibility with future uses
- If flag indicates offline keys, the offline signature section:
  Expires timestamp (4 bytes, seconds since epoch, rolls over in 2106)
  Transient sig type (2 bytes)
  Transient signing public key (length as implied by sig type)
  Signature of expires timestamp, transient sig type, and public key, by the destination public key,
  length as implied by destination public key sig type.
  This section can, and should, be generated offline.
- Data

Notes

  • Alternative is to just add a flag, and retrieve the transient public key via I2CP (See Host Lookup / Host Reply Message sections above)
  • Any other options we should add now that we have flag bytes?

SAM Changes Required

TBD. See I2CP Host Reply Message section above.

BOB Changes Required

TBD. See I2CP Host Reply Message section above.

Publishing, Migration, Compatibility

LS2 (other than encrypted LS2) is published at the same DHT location as LS1. There is no way to publish both a LS1 and LS2, unless LS2 were at a different location.

Encrypted LS2 is published at the hash of the blinded key type and key data, with daily rotation as usual.

LS2 would only be used when new features are required (new crypto, encrypted LS, meta, etc.). LS2 can only be published to floodfills of a specified version or higher.

Servers publishing LS2 would know that any connecting clients support LS2. They could send LS2 in the garlic.

Clients would send LS2 in garlics only if using new crypto. Shared clients would use LS1 indefinitely? TODO: How to have a shared clients that supports both old and new crypto?

Acknowledgements

The encrypted LS2 design is heavily influenced by Tor's v3 hidden service descriptors, which had similar design goals [TOR-REND-SPEC-V3].