Filename: 271-another-guard-selection.txt
Title: Another algorithm for guard selection
Author: Isis Lovecruft, George Kadianakis, Ola Bini, Nick Mathewson
Created: 2016-07-11
Supersedes: 259, 268
Status: Open
0.0. Preliminaries
This proposal derives from proposals 259 and 268; it is meant to
supersede both. It is in part a restatement of it, in part a
simplification, and in part a refactoring so that it does not
have the serialization problems noted by George Kadianakis. It
makes other numerous small changes. Isis, George, and Ola should
all get the credit for the well-considered ideas.
Whenever I say "Y is a subset of X" you can think in terms of
"Y-membership is a flag that can be set on members of X" or
"Y-membership is a predicate that can be evaluated on members of
X."
"More work is needed." There's a to-do at the end of the
document.
0.1. Notation: identifiers
We mention identifiers of these kinds:
[SECTIONS]
{INPUTS}, {PERSISTENT_DATA}, and {OPERATING_PARAMETERS}.
{non_persistent_data}
<states>.
Each named identifier receives a type where it is defined, and
is used by reference later on.
I'm using this convention to make it easier to tell for certain
whether every thingy we define is used, and vice versa.
1. Introduction and motivation
Tor uses entry guards to prevent an attacker who controls some
fraction of the network from observing a fraction of every user's
traffic. If users chose their entries and exits uniformly at
random from the list of servers every time they build a circuit,
then an adversary who had (k/N) of the network would deanonymize
F=(k/N)^2 of all circuits... and after a given user had built C
circuits, the attacker would see them at least once with
probability 1-(1-F)^C. With large C, the attacker would get a
sample of every user's traffic with probability 1.
To prevent this from happening, Tor clients choose a small number
of guard nodes (currently 3). These guard nodes are the only
nodes that the client will connect to directly. If they are not
compromised, the user's paths are not compromised.
But attacks remain. Consider an attacker who can run a firewall
between a target user and the Tor network, and make many of the
guards they don't control appear to be unreachable. Or consider
an attacker who can identify a user's guards, and mount
denial-of-service attacks on them until the user picks a guard
that the attacker controls.
In the presence of these attacks, we can't continue to connect to
the Tor network unconditionally. Doing so would eventually result
in the user choosing a hostile node as their guard, and losing
anonymity.
This proposal outlines a new entry guard selection algorithm,
which tries to meet the following goals:
- Heuristics and algorithms for determining how and which guards
are chosen should be kept as simple and easy to understand as
possible.
- Clients in censored regions or who are behind a fascist
firewall who connect to the Tor network should not experience
any significant disadvantage in terms of reachability or
usability.
- Tor should make a best attempt at discovering the most
appropriate behaviour, with as little user input and
configuration as possible.
- Tor clients should discover usable guards without too much
delay.
- Tor clients should resist (to the extent possible) attacks
that try to force them onto compromised guards.
2. State instances
In the algorithm below, we describe a set of persistent and
non-persistent state variables. These variables should be
treated as an object, of which multiple instances can exist.
In particular, we specify the use of three particular instances:
A. UseBridges
If UseBridges is set, then we replace the {GUARDS} set in
[Sec:GUARDS] below with the list of list of configured
bridges. We maintain a separate persistent instance of
{SAMPLED_GUARDS} and {CONFIRMED_GUARDS} and other derived
values for the UseBridges case.
B. EntryNodes / ExcludeNodes / Reachable*Addresses /
FascistFirewall / ClientUseIPv4=0
If one of the above options is set, and UseBridges is not,
then we compare the fraction of usable guards in the consensus
to the total number of guards in the consensus.
If this fraction is less than {MEANINGFUL_RESTRICTION_FRAC},
we use a separate instance of the state.
If this fraction is less than {EXTREME_RESTRICTION_FRAC}, we use a
separate instance of the state, and warn the user.
[TODO: should we have a different instance for each set of heavily
restricted options?]
C. Default
If neither of the above variant-state instances is used,
we use a default instance.
3. The algorithm.
3.0. The guards listed in the current consensus. [Section:GUARDS]
By {set:GUARDS} we mean the set of all guards in the current
consensus that are usable for all circuits. (They must have the
flags: Stable, Fast, V2Dir, Guard.)
**Rationale**
We require all guards to have the flags that we potentially need
from any guard, so that all guards are usable for all circuits.
3.1. The Sampled Guard Set. [Section:SAMPLED]
We maintain a set, {set:SAMPLED_GUARDS}, that persists across
invocations of Tor. It is an unordered subset of the nodes that
we have seen listed as a guard in the consensus at some point.
For each such guard, we record persistently:
- {pvar:ADDED_ON_DATE}: The date on which it was added to
sampled_guards.
We base this value on RAND(now, {GUARD_LIFETIME}/10). See
Appendix [RANDOM] below.
- {pvar:ADDED_BY_VERSION}: The version of Tor that added it to
sampled_guards.
- {pvar:IS_LISTED}: Whether it was listed as a usable Guard in
the _most recent_ consensus we have seen.
- {pvar:FIRST_UNLISTED_AT}: If IS_LISTED is false, the publication date
of the earliest consensus in which this guard was listed such that we
have not seen it listed in any later consensus. Otherwise "None."
We randomize this, based on
RAND(added_at_time, {REMOVE_UNLISTED_GUARDS_AFTER} / 5)
For each guard in {SAMPLED_GUARDS}, we also record this data,
non-persistently:
- {tvar:last_tried_connect}: A 'last tried to connect at'
time. Default 'never'.
- {tvar:is_reachable}: an "is reachable" tristate, with
possible values { <state:yes>, <state:no>, <state:maybe> }.
Default '<maybe>.'
[Note: "yes" is not strictly necessary, but I'm
making it distinct from "maybe" anyway, to make our
logic clearer. A guard is "maybe" reachable if it's
worth trying. A guard is "yes" reachable if we tried
it and succeeded.]
- {tvar:failing_since}: The first time when we failed to
connect to this guard. Defaults to "never". Reset to
"never" when we successfully connect to this guard.
- {tvar:is_pending} A "pending" flag. This indicates that we
are trying to build an exploratory circuit through the
guard, and we don't know whether it will succeed.
We require that {SAMPLED_GUARDS} contain at least
{MIN_SAMPLE_THRESHOLD} of the number of guards in the consensus
(if possible), but not more than {MAX_SAMPLE_THRESHOLD} of the
number of guards in the consensus.
To add a new guard to {SAMPLED_GUARDS}, pick an entry at random
from ({GUARDS} - {SAMPLED_GUARDS}), weighted by bandwidth.
We remove an entry from {SAMPLED_GUARDS} if:
* We have a live consensus, and {IS_LISTED} is false, and
{FIRST_UNLISTED_AT} is over {REMOVE_UNLISTED_GUARDS_AFTER}
days in the past.
OR
* We have a live consensus, and we cannot parse
{ADDED_BY_VERSION}.
OR
* We have a live consensus, and {ADDED_ON_DATE} is over
{GUARD_LIFETIME} ago, *and* {CONFIRMED_ON_DATE} is either
"never", or over {GUARD_CONFIRMED_MIN_LIFETIME} ago.
Note that {SAMPLED_GUARDS} does not depend on our configuration.
It is possible that we can't actually connect to any of these
guards.
**Rationale**
The {SAMPLED_GUARDS} set is meant to limit the total number of
guards that a client will connect to in a given period. The
upper limit on its size prevents us from considering too many
guards.
The first expiration mechanism is there so that our
{SAMPLED_GUARDS} list does not accumulate so many dead
guards that we cannot add new ones.
The second expiration mechanism makes us rotate our guards slowly
over time.
3.2. The Usable Sample [Section:FILTERED]
We maintain another set, {set:FILTERED_GUARDS}, that does not
persist. It is derived from:
- {SAMPLED_GUARDS}
- our current configuration,
- the path bias information.
A guard is a member of {set:FILTERED_GUARDS} if and only if all
of the following are true:
- It is a member of {SAMPLED_GUARDS}, with {IS_LISTED} set to
true.
- It is not disabled because of path bias issues.
- It is not disabled because of ReachableAddress police,
the ClientUseIPv4 setting, the ClientUseIPv6 setting,
the FascistFirewall setting, or some other
option that prevents using some addresses.
- It is not disabled because of ExcludeNodes.
- It is a bridge if UseBridges is true; or it is not a
bridge if UseBridges is false.
We have an additional subset, {set:USABLE_FILTERED_GUARDS}, which
is defined to be the subset of {FILTERED_GUARDS} where
{is_reachable} is <yes> or <maybe>.
We try to maintain a requirement that {USABLE_FILTERED_GUARDS}
contain at least {MIN_FILTERED_SAMPLE} elements:
Whenever we are going to sample from {USABLE_FILTERED_GUARDS},
and it contains fewer than {MIN_FILTERED_SAMPLE} elements, we
add new elements to {SAMPLED_GUARDS} until one of the following
is true:
* {USABLE_FILTERED_GUARDS} is large enough,
OR
* {SAMPLED_GUARDS} is at its maximum size.
** Rationale **
These filters are applied _after_ sampling: if we applied them
before the sampling, then our sample would reflect the set of
filtering restrictions that we had in the past.
3.3. The confirmed-guard list. [Section:CONFIRMED]
[formerly USED_GUARDS]
We maintain a persistent ordered list, {list:CONFIRMED_GUARDS}.
It contains guards that we have used before, in our preference
order of using them. It is a subset of {SAMPLED_GUARDS}. For
each guard in this list, we store persistently:
- {pvar:IDENTITY} Its fingerprint
- {pvar:CONFIRMED_ON_DATE} When we added this guard to
{CONFIRMED_GUARDS}.
Randomized as RAND(now, {GUARD_LIFETIME}/10).
We add new members to {CONFIRMED_GUARDS} when we mark a circuit
built through a guard as "for user traffic."
Whenever we remove a member from {SAMPLED_GUARDS}, we also remove
it from {CONFIRMED_GUARDS}.
[Note: You can also regard the {CONFIRMED_GUARDS} list as a
total ordering defined over a subset of {SAMPLED_GUARDS}.]
Definition: we call Guard A "higher priority" than another Guard B
if, when A and B are both reachable, we would rather use A. We
define prioirty as follows:
* Every guard in {CONFIRMED_GUARDS} has a higher priority
than every guard not in {CONFIRMED_GUARDS}.
* Among guards in {CONFIRMED_GUARDS}, the one appearing earlier
on the {CONFIRMED_GUARDS} list has a higher priority.
* Among guards that do not appear in {CONFIRMED_GUARDS},
{is_pending}==true guards have higher priority.
* Among those, the guard with earlier {last_tried_connect} time
have higher priority.
* Finally, among guards that do not appear in
{CONFIRMED_GUARDS} with {is_pending==false}, all have equal
priority.
** Rationale **
We add elements to this ordering when we have actually used them
for building a usable circuit. We could mark them at some other
time (such as when we attempt to connect to them, or when we
actually connect to them), but this approach keeps us from
committing to a guard before we actually use it for sensitive
traffic.
3.4. The Primary guards [Section:PRIMARY]
We keep a run-time non-persistent ordered list of
{list:PRIMARY_GUARDS}. It is a subset of {FILTERED_GUARDS}. It
contains {N_PRIMARY_GUARDS} elements.
To compute primary guards, take the ordered intersection of
{CONFIRMED_GUARDS} and {FILTERED_GUARDS}, and take the first
{N_PRIMARY_GUARDS} elements. If there are fewer than
{N_PRIMARY_GUARDS} elements, add additional elements to
PRIMARY_GUARDS chosen _uniformly_ at random from
({FILTERED_GUARDS} - {CONFIRMED_GUARDS}).
Note that {PRIMARY_GUARDS} do not have to be in
{USABLE_FILTERED_GUARDS}: they might be unreachable.
** Rationale **
These guards are treated differently from other guards. If one of
them is usable, then we use it right away. For other guards
{FILTERED_GUARDS}, if it's usable, then before using it we might
first double-check whether perhaps one of the primary guards is
usable after all.
3.5. Retrying guards. [Section:RETRYING]
(We run this process as frequently as needed. It can be done once
a second, or just-in-time.)
If a primary sampled guard's {is_reachable} status is <no>, then
we decide whether to update its {is_reachable} status to <maybe>
based on its {last_tried_connect} time, its {failing_since} time,
and the {PRIMARY_GUARDS_RETRY_SCHED} schedule.
If a non-primary sampled guard's {is_reachable} status is <no>, then
we decide whether to update its {is_reachable} status to <maybe>
based on its {last_tried_connect} time, its {failing_since} time,
and the {GUARDS_RETRY_SCHED} schedule.
** Rationale **
An observation that a guard has been 'unreachable' only lasts for
a given amount of time, since we can't infer that it's unreachable
now from the fact that it was unreachable a few minutes ago.
3.6. Selecting guards for circuits. [Section:SELECTING]
Every origin circuit is now in one of these states:
<state:usable_on_completion>,
<state:usable_if_no_better_guard>,
<state:waiting_for_better_guard>, or
<state:complete>.
You may only attach streams to <complete> circuits.
(Additionally, you may only send RENDEZVOUS cells, ESTABLISH_INTRO
cells, and INTRODUCE cells on <complete> circuits.)
The per-circuit state machine is:
New circuits are <usable_on_completion> or
<usable_if_no_better_guard>.
A <usable_on_completion> circuit may become <complete>, or may
fail.
A <usable_if_no_better_guard> circuit may become
<usable_on_completion>; may become <waiting_for_better_guard>; or may
fail.
A <waiting_for_better_guard> circuit will become <complete>, or will
be closed, or will fail.
A <complete> circuit remains <complete> until it fails or is
closed.
Each of these transitions is described below.
We keep, as global transient state:
* {tvar:last_time_on_internet} -- the last time at which we
successfully used a circuit or connected to a guard. At
startup we set this to "infinitely far in the past."
When we want to build a circuit, and we need to pick a guard:
* If any entry in PRIMARY_GUARDS has {is_reachable} status of
<maybe> or <yes>, return the first such guard. The circuit is
<usable_on_completion>.
[Note: We do not use {is_pending} on primary guards, since we
are willing to try to build multiple circuits through them
before we know for sure whether they work, and since we will
not use any non-primary guards until we are sure that the
primary guards are all down. (XX is this good?)]
* Otherwise, if the ordered intersection of {CONFIRMED_GUARDS}
and {USABLE_FILTERED_GUARDS} is nonempty, return the first
entry in that intersection that has {is_pending} set to
false. Set its value of {is_pending} to true. The circuit
is now <usable_if_no_better_guard>. (If all entries have
{is_pending} true, pick the first one.)
* Otherwise, if there is no such entry, select a member at
random from {USABLE_FILTERED_GUARDS}. Set its {is_pending}
field to true. The circuit is <usable_if_no_better_guard>.
We update the {last_tried_connect} time for the guard to 'now.'
** Rationale **
We're getting to the core of the algorithm here. Our main goals are to
make sure that
1. If it's possible to use a primary guard, we do.
2. We probably use the first primary guard.
So we only try non-primary guards if we're pretty sure that all
the primary guards are down, and we only try a given primary guard
if the earlier primary guards seem down.
When we _do_ try non-primary guards, however, we only build one
circuit through each, to give it a chance to succeed or fail. If
ever such a circuit succeeds, we don't use it until we're pretty
sure that it's the best guard we're getting. (see below).
[XXX timeout.]
3.7. When a circuit fails. [Section:ON_FAIL]
When a circuit fails in a way that makes us conclude that a guard
is not reachable, we take the following steps:
* We set the guard's {is_reachable} status to <no>. If it had
{is_pending} set to true, we make it non-pending.
* We close the circuit, of course. (This removes it from
consideration by the algorithm in [UPDATE_WAITING].)
* Update the list of waiting circuits. (See [UPDATE_WAITING]
below.)
[Note: the existing Tor logic will cause us to create more
circuits in response to some of these steps; and also see
[ON_CONSENSUS].]
** Rationale **
See [SELECTING] above for rationale.
3.8. When a circuit succeeds [Section:ON_SUCCESS]
When a circuit succeeds in a way that makes us conclude that a
guard _was_ reachable, we take these steps:
* We set its {is_reachable} status to <yes>.
* We set its {failing_since} to "never".
* If the guard was {is_pending}, we clear the {is_pending} flag.
* If the guard was not a member of {CONFIRMED_GUARDS}, we add
it to the end of {CONFIRMED_GUARDS}.
* If this circuit was <usable_on_completion>, this circuit is
now <complete>. You may attach streams to this circuit,
and use it for hidden services.
* If this circuit was <usable_if_no_better_guard>, it is now
<waiting_for retry>. You may not yet attach streams to it.
Then check whether the {last_time_on_internet} is more than
{INTERNET_LIKELY_DOWN_INTERVAL} seconds ago:
* If it is, then mark all {PRIMARY_GUARDS} as "maybe"
reachable.
* If it is not, update the list of waiting circuits. (See
[UPDATE_WAITING] below)
[Note: the existing Tor logic will cause us to create more
circuits in response to some of these steps; and see
[ON_CONSENSUS].]
** Rationale **
See [SELECTING] above for rationale.
3.9. Updating the list of waiting circuits [Section:UPDATE_WAITING]
We run this procedure whenever it's possible that a
<waiting_for_better_guard> circuit might be ready to be called
<complete>.
* If any circuit is <waiting_for_better_guard>, and every currently
{is_pending} circuit whose guard has higher priority has been
in state <usable_if_no_better_guard> for at least
{NONPRIMARY_GUARD_CONNECT_TIMEOUT} seconds, and all primary
guards have reachable status of <no>, then call that circuit
<complete>.
* If any circuit is <complete>, then do not use any
<waiting_for_better_guard> or <usable_if_no_better_guard> circuits
circuits whose guards have lower priority. (Time them out
after a {NONPRIMARY_GUARD_IDLE_TIMEOUT} seconds.)
**Rationale**
If we open a connection to a guard, we might want to use it
immediately (if we're sure that it's the best we can do), or we
might want to wait a little while to see if some other circuit
which we like better will finish.
When we mark a circuit <complete>, we don't close the
lower-priority circuits immediately: we might decide to use
them after all if the <complete> circuit goes down before
{NONPRIMARY_GUARD_IDLE_TIMEOUT} seconds.
3.10. Whenever we get a new consensus. [Section:ON_CONSENSUS]
We update {GUARDS}.
For every guard in {SAMPLED_GUARDS}, we update {IS_LISTED} and
{FIRST_UNLISTED_AT}.
[**] We remove entries from {SAMPLED_GUARDS} if appropriate,
according to the sampled-guards expiration rules. If they were
in {CONFIRMED_GUARDS}, we also remove them from
{CONFIRMED_GUARDS}.
We recompute {FILTERED_GUARDS}, and everything that derives from
it, including {USABLE_FILTERED_GUARDS}, and {PRIMARY_GUARDS}.
(Whenever one of the configuration options that affects the
filter is updated, we repeat the process above, starting at the
[**] line.)
3.11. Deciding whether to generate a new circuit.
[Section:NEW_CIRCUIT_NEEDED]
In current Tor, we generate a new circuit when we don't have
enough circuits either built or in-progress to handle a given
stream, or an expected stream.
For the purpose of this rule, we say that <waiting_for_better_guard>
circuits are neither built nor in-progress; that <complete>
circuits are built; and that the other states are in-progress.
A. Appendices
A.1. Parameters with suggested values. [Section:PARAM_VALS]
(All suggested values chosen arbitrarily)
{param:MIN_SAMPLE_THRESHOLD} -- 15
{param:MAX_SAMPLE_THRESHOLD} -- 50
{param:GUARD_LIFETIME} -- 120 days
{param:REMOVE_UNLISTED_GUARDS_AFTER} -- 20 days
[previously ENTRY_GUARD_REMOVE_AFTER]
{param:MIN_FILTERED_SAMPLE} -- 10
{param:N_PRIMARY_GUARDS} -- 3
{param:PRIMARY_GUARDS_RETRY_SCHED}
-- every 30 minutes for the first 6 hours.
-- every 2 hours for the next 3.75 days.
-- every 4 hours for the next 3 days.
-- every 9 hours thereafter.
{param:GUARDS_RETRY_SCHED} -- 1 hour
-- every hour for the first 6 hours.
-- every 4 hours for the next 3.75 days.
-- every 18 hours for the next 3 days.
-- every 36 hours thereafter.
{param:INTERNET_LIKELY_DOWN_INTERVAL} -- 10 minutes
{param:NONPRIMARY_GUARD_CONNECT_TIMEOUT} -- 15 seconds
{param:NONPRIMARY_GUARD_IDLE_TIMEOUT} -- 10 minutes
{param:MEANINGFUL_RESTRICTION_FRAC} -- .2
{param:EXTREME_RESTRICTION_FRAC} -- .01
{param:GUARD_CONFIRMED_MIN_LIFETIME} -- 60 days
A.2. Random values [Section:RANDOM]
Frequently, we want to randomize the expiration time of something
so that it's not easy for an observer to match it to its start
time. We do this by randomizing its start date a little, so that
we only need to remember a fixed expiration interval.
By RAND(now, INTERVAL) we mean a time between now and INTERVAL in
the past, chosen uniformly at random.
A.3. Why not a sliding scale of primaryness? [Section:CVP]
At one meeting, I floated the idea of having "primaryness" be a
continuous variable rather than a boolean.
I'm no longer sure this is a great idea, but I'll try to outline
how it might work.
To begin with: being "primary" gives it a few different traits:
1) We retry primary guards more frequently. [Section:RETRYING]
2) We don't even _try_ building circuits through
lower-priority guards until we're pretty sure that the
higher-priority primary guards are down. (With non-primary
guards, on the other hand, we launch exploratory circuits
which we plan not to use if higher-priority guards
succeed.) [Section:SELECTING]
3) We retry them all one more time if a circuit succeeds after
the net has been down for a while. [Section:ON_SUCCESS]
We could make each of the above traits continuous:
1) We could make the interval at which a guard is retried
depend continuously on its position in CONFIRMED_GUARDS.
2) We could change the number of guards we test in parallel
based on their position in CONFIRMED_GUARDS.
3) We could change the rule for how long the higher-priority
guards need to have been down before we call a
<usable_if_no_better_guard> circuit <complete> based on a
possible network-down condition. For example, we could
retry the first guard if we tried it more than 10 seconds
ago, the second if we tried it more than 20 seconds ago,
etc.
I am pretty sure, however, that if these are worth doing, they
need more analysis! Here's why:
* They all have the potential to leak more information about a
guard's exact position on the list. Is that safe? Is there
any way to exploit that? I don't think we know.
* They all seem like changes which it would be relatively
simple to make to the code after we implement the simpler
version of the algorithm described above.
TODO. Still non-addressed issues [Section:TODO]
Formats to use when making information persistent
Migration from old data format to new.
Explain the overall flow of the circuit creation and guard
picking algorithms, if they are not clear.
Simulate to answer: Will this work in a dystopic world?
Simulate actual behavior.
For all lifetimes: instead of storing the "this began at" time,
store the "remove this at" time, slightly randomized.
Clarify that when you get a <complete> circuit, you might need to
relaunch circuits through that same guard immediately, if they
are circuits that have to be independent.
Fix all items marked XX or TODO.
"Directory guards" -- do they matter?
Suggestion: require that all guards support downloads via BEGINDIR.
We don't need to worry about directory guards for relays, since we
aren't trying to prevent relay enumeration.
IP version preferenes via ClientPreferIPv6ORPort
Suggestion: Treat it as a preference when adding to
{CONFIRMED_GUARDS}, but not otherwise.
