Filename: 300-walking-onions.txt
Title: Walking Onions: Scaling and Saving Bandwidth
Author: Nick Mathewson
Created: 5-Feb-2019
Status: Informational


0. Status

   This proposal describes a mechanism called "Walking Onions" for
   scaling the Tor network and reducing the amount of client bandwidth
   used to maintain a client's view of the Tor network.

   This is a draft proposal; there are problems left to be solved and
   questions left to be answered.  Proposal 323 tries to fill in all the
   gaps.

1. Introduction

   In the current Tor network design, we assume that every client has a
   complete view of all the relays in the network.  To achieve this,
   clients download consensus directories at regular intervals, and
   download descriptors for every relay listed in the directory.

   The substitution of microdescriptors for regular descriptors
   (proposal 158) and the use of consensus diffs (proposal 140) have
   lowered the bytes that clients must dedicate to directory operations.
   But we still face the problem that, if we force each client to know
   about every relay in the network, each client's directory traffic
   will grow linearly with the number of relays in the network.

   Another drawback in our current system is that client directory
   traffic is front-loaded: clients need to fetch an entire directory
   before they begin building circuits.  This places extra delays on
   clients, and extra load on the network.

   To anonymize the world, we will need to scale to a much larger number
   of relays and clients: requiring clients to know about every relay in
   the set simply won't scale, and requiring every new client to download
   a large document is also problematic.

   There are obvious responses here, and some other anonymity tools have
   taken them.  It's possible to have a client only use a fraction of
   the relays in a network--but doing so opens the client to _epistemic
   attacks_, in which the difference in clients' views of the
   network is used to partition their traffic.  It's also possible to
   move the problem of selecting relays from the client to the relays
   themselves, and let each relay select the next relay in turn--but
   this choice opens the client to _route capture attacks_, in which a
   malicious relay selects only other malicious relays.

   In this proposal, I'll describe a design for eliminating up-front
   client directory downloads.  Clients still choose relays at random,
   but without ever having to hold a list of all the relays. This design
   does not require clients to trust relays any more than they do today,
   or open clients to epistemic attacks.

   I hope to maintain feature parity with the current Tor design; I'll
   list the places in which I haven't figured out how to do so yet.

   I'm naming this design "walking onions".  The walking onion (Allium x
   proliferum) reproduces by growing tiny little bulbs at the
   end of a long stalk.  When the stalk gets too top-heavy, it flops
   over, and the little bulbs start growing somewhere new.

   The rest of this document will run as follows.  In section 2, I'll
   explain the ideas behind the "walking onions" design, and how they
   can eliminate the need for regular directory downloads.  In section 3, I'll
   answer a number of follow-up questions that arise, and explain how to
   keep various features in Tor working.  Section 4 (not yet written)
   will elaborate all the details needed to turn this proposal into a
   concrete set of specification changes.

2. Overview

2.1. Recapping proposal 141

   Back in Proposal 141 ("Download server descriptors on demand"), Peter
   Palfrader proposed an idea for eliminating ahead-of-time descriptor
   downloads.  Instead of fetching all the descriptors in advance, a
   client would fetch the descriptor for each relay in its path right
   before extending the circuit to that relay.  For example, if a client
   has a circuit from A->B and wants to extend the circuit to C, the
   client asks B for C's descriptor, and then extends the circuit to C.

   (Note that the client needs to fetch the descriptor every time it
   extends the circuit, so that an observer can't tell whether the
   client already had the descriptor or not.)

   There are a couple of limitations for this design:
      * It still requires clients to download a consensus.
      * It introduces a extra round-trip to each hop of the circuit
        extension process.

   I'll show how to solve these problems in the two sections below.

2.2. An observation about the ntor handshake.

   I'll start with an observation about our current circuit extension
   handshake, ntor: it should not actually be necessary to know a
   relay's onion key before extending to it.

   Right now, the client sends:
         NODEID     (The relay's identity)
         KEYID      (The relay's public onion key)
         CLIENT_PK  (a diffie-hellman public key)

   and the relay responds with:
         SERVER_PK  (a diffie-hellman public key)
         AUTH       (a function of the relay's private keys and
                     *all* of the public keys.)

   Both parties generate shared symmetric keys from the same inputs
   that are are used to create the AUTH value.

   The important insight here is that we could easily change
   this handshake so that the client sends only CLIENT_PK, and receives
   NODEID and KEYID as part of the response.

   In other words, the client needs to know the relay's onion key to
   _complete_ the handshake, but doesn't actually need to know the
   relay's onion key in order to _initiate_ the handshake.

   This is the insight that will let us save a round trip:  When the
   client goes to extend a circuit from A->B to C, it can send B a
   request to extend to C and retrieve C's descriptor in a single step.
   Specifically, the client sends only CLIENT_PK, and relay B can include C's
   keys as part of the EXTENDED cell.

2.3. Extending by certified index

   Now I'll explain how the client can avoid having to download a
   list of relays entirely.

   First, let's look at how a client chooses a random relay today.
   First, the client puts all of the relays in a list, and computes a
   weighted bandwidth for each one. For example, suppose the relay
   identities are R1, R2, R3, R4, and R5, and their bandwidth weights
   are 50, 40, 30, 20, and 10.  The client makes a table like this:

      Relay   Weight     Range of index values
      R1      50         0..49
      R2      40         50..89
      R3      30         90..119
      R4      20         120..139
      R5      10         140..149

   To choose a random relay, the client picks a random index value
   between 0 and 149 inclusive, and looks up the corresponding relay in
   the table.  For example, if the client's random number is 77, it will
   choose R2.  If its random number is 137, it chooses R4.

   The key observation for the "walking onions" design is that the
   client doesn't actually need to construct this table itself.
   Instead, we will have this table be constructed by the authorities
   and distributed to all the relays.

   Here's how it works: let's have the authorities make a new kind of
   consensus-like thing.  We'll call it an Efficient Network Directory
   with Individually Verifiable Entries, or "ENDIVE" for short.  This
   will differ from the client's index table above in two ways.  First,
   every entry in the ENDIVE is normalized so that the bandwidth
   weights maximum index is 2^32-1:

       Relay      Normalized weight    Range of index values
       R1         0x55555546           0x00000000..0x55555545
       R2         0x44444438           0x55555546..0x9999997d
       R3         0x3333332a           0x9999997e..0xcccccca7
       R4         0x2222221c           0xcccccca8..0xeeeeeec3
       R5         0x1111113c           0xeeeeeec4..0xffffffff

   Second, every entry in the ENDIVE is timestamped and signed by the
   authorities independently, so that when a client sees a line from the
   table above, it can verify that it came from an authentic ENDIVE.
   When a client has chosen a random index, one of these entries will
   prove to the client that a given relay corresponds to that index.
   Because of this property, we'll be calling these entries "Separable
   Network Index Proofs", or "SNIP"s for short.

   For example, a single SNIP from the table above might consist of:
     * A range of times during which this SNIP is valid
     * R1's identity
     * R1's ntor onion key
     * R1's address
     * The index range 0x00000000..0x55555545
     * A signature of all of the above, by a number of authorities

   Let's put it together. Suppose that the client has a circuit from
   A->B, and it wants to extend to a random relay, chosen randomly
   weighted by bandwidth.

   1. The client picks a random index value between 0 and 2**32 - 1.  It
      sends that index to relay B in its EXTEND cell, along with a
      g^x value for the ntor handshake.

      Note: the client doesn't send an address or identity for the next
      relay, since it doesn't know what relay it has chosen!  (The
      combination of an index and a g^x value is what I'm calling a
      "walking onion".)

   2. Now, relay B looks up the index in its most recent ENDIVE, to
      learn which relay the client selected.

      (For example, suppose that the client's random index value is
      0x50000001.  This index value falls between 0x00000000 and
      0x55555546 in the table above, so the relay B sees that the client
      has chosen R1 as its next hop.)

   3. Relay B sends a create cell to R1 as usual.  When it gets a
      CREATED reply, it includes the authority-signed SNIP for
      R1 as part of the EXTENDED cell.

   4. As part of verifying the handshake, the client verifies that the
      SNIP was signed by enough authorities, that its timestamp
      is recent enough, and that it actually corresponds to the
      random index that the client selected.

   Notice the properties we have with this design:

       - Clients can extend circuits without having a list of all the
         relays.

       - Because the client's random index needs to match a routing
         entry signed by the authorities, the client is still selecting
         a relay randomly by weight.  A hostile relay cannot choose
         which relay to send the client.


   On a failure to extend, a relay should still report the routing entry
   for the other relay that it couldn't connect to.  As before, a client
   will start a new circuit if a partially constructed circuit is a
   partial failure.


   We could achieve a reliability/security tradeoff by letting clients
   offer the relay a choice of two or more indices to extend to.
   This would help reliability, but give the relay more influence over
   the path.  We'd need to analyze this impact.


   In the next section, I'll discuss a bunch of details that we need to
   straighten out in order to make this design work.


3. Sorting out the details.

3.1. Will these routing entries fit in EXTEND2 and EXTENDED2 cells?

   The EXTEND2 cell is probably big enough for this design.  The random
   index that the client sends can be a new "link specifier" type,
   replacing the IP and identity link specifiers.

   The EXTENDED2 cell is likely to need to grow here.  We'll need to
   implement proposal 249 ("Allow CREATE cells with >505 bytes of
   handshake data") so that EXTEND2 and EXTENDED2 cells can be larger.

3.2. How should SNIPs be signed?

   We have a few options, and I'd like to look into the possibilities
   here more closely.

   The simplest possibility is to use **multiple signatures** on each
   SNIP, the way we do today for consensuses.  These signatures should
   be made using medium-term Ed25519 keys from the authorities.  At a
   cost of 64 bytes per signature, at 9 authorities, we would need 576
   bytes for each SNIP.  These signatures could be batch-verified to
   save time at the client side.  Since generating a signature takes
   around 20 usec on my mediocre laptop, authorities should be able to
   generate this many signatures fairly easily.

   Another possibility is to use a **threshold signature** on each SNIP,
   so that the authorities collaboratively generate a short signature
   that the clients can verify.  There are multiple threshold signature
   schemes that we could consider here, though I haven't yet found one
   that looks perfect.

   Another possibility is to use organize the SNIPs in a **merkle tree
   with a signed root**.  For this design, clients could download the
   signed root periodically, and receive the hash-path from the signed
   root to the SNIP.  This design might help with
   certificate-transparency-style designs, and it would be necessary if we
   ever want to move to a postquantum signature algorithm that requires
   large signatures.

   Another possibility (due to a conversation among Chelsea Komlo, Sajin
   Sasy, and Ian Goldberg), is to *use SNARKs*.  (Why not?  All the cool
   kids are doing it!)  For this, we'd have the clients download a
   signed hash of the ENDIVE periodically, and have the authorities
   generate a SNARK for each SNIP, proving its presence in that
   document.

3.3. How can we detect authority misbehavior?

   We might want to take countermeasures against the possibility that a
   quorum of corrupt or compromised authorities give some relays a
   different set of SNIPs than they give other relays.

   If we incorporate a merkle tree or a hash chain in the design, we can
   use mechanisms similar to certificate transparency to ensure that the
   authorities have a consistent log of all the entries that they have
   ever handed out.

3.4. How many types of weighted node selection are there, and how do we
     handle them?

   Right now, there are multiple weights that we use in Tor:
      * Weight for exit
      * Weight for guard
      * Weight for middle node

   We also filter nodes for several properties, such as flags they have.

   To reproduce this behavior, we should enumerate the various weights
   and filters that we use, and (if there are not too many) create a
   separate index for each.  For example, the Guard index would weight
   every node for selection as guard, assigning 0 weight to non-Guard
   nodes.  The Exit index would weight every node for selection as an
   exit, assigning 0 weight to non-Exit nodes.

   When choosing a relay, the client would have to specify which index
   to use.  We could either have a separate (labeled) set of SNIPs
   entries for each index, or we could have each SNIP have a separate
   (labeled) index range for each index.

   REGRESSION: the client's choice of which index to use would leak the
   next router's position and purpose in the circuit.  This information
   is something that we believe relays can infer now, but it's not a
   desired feature that they can.

3.5. Does this design break onion service introduce handshakes?

   In rend-spec-v3.txt section 3.3.2, we specify a variant of ntor for
   use in INTRODUCE2 handshakes.  It allows the client to send encrypted
   data as part of its initial ntor handshake, but requires the client
   to know the onion service's onion key before it sends its initial
   handshake.

   That won't be a problem for us here, though: we still require clients
   to fetch onion service descriptors before contacting a onion
   service.

3.6. How does the onion service directory work here?

   The onion service directory is implemented as a hash ring, where
   each relay's position in the hash ring is decided by a hash of its
   identity, the current date, and a shared random value that the
   authorities compute each day.

   To implement this hash ring using walking onions, we would need to
   have an extra index based not on bandwidth, but on position in the
   hash ring.  Then onion services and clients could build a circuit,
   then extend it one more hop specifying their desired index in the
   hash ring.

   We could either have a command to retrieve a trio of hashring-based
   routing entries by index, or to retrieve (or connect to?) the n'th item
   after a given hashring entry.

3.7. How can clients choose guard nodes?

   We can reuse the fallback directories here.  A newly bootstrapping
   client would connect to a fallback directory, then build a three-hop
   circuit, and finally extend the three-hop circuit by indexing to a
   random guard node.  The random guard node's SNIP would
   contain the information that the client needs to build real circuits
   through that guard in the future.  Because the client would be
   building a three-hop circuit, the fallback directory would not learn
   the client's guards.

   (Note that even if the extend attempt fails, we should still pick the
   node as a possible guard based on its router entry, so that other
   nodes can't veto our choice of guards.)

3.8. Does the walking onions design preclude postquantum circuit handshakes?

   Not at all!  Both proposal 263 (ntru) and proposal 270 (newhope) work
   by having the client generate an ephemeral key as part of its initial
   handshake.  The client does not need to know the relay's onion key to
   do this, so we can still integrate those proposals with this one.

3.9. Does the walking onions design stop us from changing the network
     topology?

   For Tor to continue to scale, we will someday need to accept that not
   every relay can be simultaneously connected to every other relay.
   Therefore, we will need to move from our current clique topology
   assumption to some other topology.

   There are also proposals to change node selection rules to generate
   routes providing better performance, or improved resistance to local
   adversaries.

   We can, I think, implement this kind of proposal by changing the way
   that ENDIVEs are generated.  Instead giving every relay the same
   ENDIVE, the authorities would generate different ENDIVEs for
   different relays, depending on the probability distribution of which
   relay should be chosen after which in the network topology.  In the
   extreme case, this would produce O(n) ENDIVEs and O(n^2) SNIPs.  In
   practice, I hope that we could do better by having the network
   topology be non-clique, and by having many relays share the same
   distribution of successors.


3.10. How can clients handle exit policies?

   This is an unsolved challenge.  If the client tells the middle relay
   its target port, it leaks information inappropriately.

   One possibility is to try to gather exit policies into common
   categories, such as "most ports supported" and "most common ports
   supported".

   Another (inefficient) possibility is for clients to keep trying exits
   until they find one that works.

   Another (inefficient) possibility is to require that clients who use
   unusual ports fall back to the old mechanism for route selection.


3.11. Can this approach support families?

   This is an unsolved challenge.

   One (inefficient) possibility is for clients to generate circuits and
   discard those that use multiple relays in the same family.

   One (not quite compatible) possibility is for the authorities to sort
   the ENDIVE so that relays in the same family are adjacent to
   one another.  The index-bounds part of each SNIP would also
   have to include the bounds of the family.  This approach is not quite
   compatible with the status quo, because it prevents relays from
   belonging to more than one family.

   One interesting possibility (due to Chelsea Komlo, Sajin Sasy, and
   Ian Goldberg) is for the middle node to take responsibility for
   family enforcement. In this design, the client might offer the middle
   node multiple options for the next relay's index, and the middle node
   would choose the first such relay that is neither in its family nor
   its predecessor's family.  We'd need to look for a way to make sure
   that the middle node wasn't biasing the path selection.

   (TODO: come up with more ideas here.)

3.12. Can walking onions support IP-based and country-based restrictions?

   This is an unsolved challenge.

   If the user's restrictions do not exclude most paths, one
   (inefficient) possibility is for the user to generate paths until
   they generate one that they like.  This idea becomes inefficient
   if the user is excluding most paths.

   Another (inefficient and fingerprintable) possibility is to require
   that clients who use complex path restrictions fall back to the old
   mechanism for route selection.

   (TODO: come up with better ideas here.)

3.13. What scaling problems have we not solved with this design?

   The walking onions design doesn't solve (on its own) the problem that
   the authorities need to know about every relay, and arrange to have
   every relay tested.

   The walking onions design doesn't solve (on its own) the problem that
   relays need to have a list of all the relays.  (But see section 3.9
   above.)

3.14. Should we still have clients download a consensus when they're
      using walking onions?

   There are some fields in the current consensus directory documents
   that the clients will still need, like the list of supported
   protocols and network parameters.  A client that uses walking onions
   should download a new flavor of consensus document that contains only
   these fields, and does not list any relays.  In some signature
   schemes, this consensus would contain a digest of the ENDIVE -- see
   3.2 above.

   (Note that this document would be a "consensus document" but not a
   "consensus directory", since it doesn't list any relays.)


4. Putting it all together

   [This is the section where, in a later version of this proposal, I
   would specify the exact behavior and data formats to be used here.
   Right now, I'd say we're too early in the design phase.]


A.1. Acknowledgments

   Thanks to Peter Palfrader for his original design in proposal 141,
   and to the designers of PIR-Tor, both of which inspired aspects of
   this Walking Onions design.

   Thanks to Chelsea Komlo, Sajin Sasy, and Ian Goldberg for feedback on
   an earlier version of this design.

   Thanks to David Goulet, Teor, and George Kadianakis for commentary on
   earlier versions of this draft.

   This research was supported by NSF grants CNS-1526306 and CNS-1619454.

A.2. Additional ideas

   Teor notes that there are ways to try to get this idea to apply to
   one-pass circuit construction, something like the old onion design.
   We might be able to derive indices and keys from the same seeds,
   even.  I don't see a way to do this without losing forward secrecy,
   but it might be worth looking at harder.