Filename: 254-padding-negotiation.txt
Title: Padding Negotiation
Authors: Mike Perry
Created: 01 September 2015
Status: Closed

   [See padding-spec.txt for the implemented version of this proposal.]

0. Overview

This proposal aims to describe mechanisms for requesting various types
of padding from relays.

These padding primitives are general enough to use to defend against
both website traffic fingerprinting as well as hidden service circuit
setup fingerprinting.

1. Motivation

Tor already supports both link-level padding via (CELL_PADDING cell
types), as well as circuit-level padding (via RELAY_COMMAND_DROP relay

Unfortunately, there is no way for clients to request padding from
relays, or request that relays not send them padding to conserve
bandwidth. This proposal aims to create a mechanism for clients to do
both of these.

It also establishes consensus parameters to limit the amount of padding
that relays will send, to prevent custom wingnut clients from requesting
too much.

2. Link-level padding

Padding is most useful if it can defend against a malicious or
compromised guard node. However, link-level padding is still useful to
defend against an adversary that can merely observe a Guard node
externally, such as for low-resolution netflow-based attacks (see
Proposal 251[1]).

In that scenario, the primary negotiation mechanism we need is a way for
mobile clients to tell their Guards to stop padding, or to pad less
often. The following Trunnel payload should cover the needed


  /* The start command tells the relay to alter its min and max netflow
     timeout range values, and send padding at that rate (resuming
     if stopped). The stop command tells the relay to stop sending
     link-level padding. */
  struct channelpadding_negotiate {
    u8 version IN [0];

    /* Min must not be lower than the current consensus parameter
       nf_ito_low. Ignored if command is stop. */
    u16 ito_low_ms;

    /* Max must not be lower than ito_low_ms. Ignored if command is stop. */
    u16 ito_high_ms;

After the above cell is received, the guard should use the 'ito_low_ms' and
'ito_high_ms' values as the minimum and maximum values (respectively) for
inactivity before it decides to pad the channel. The actual timeout value is
randomly chosen between those two values through an appropriate probability
distribution (see proposal251 for the netflow padding protocol).

More complicated forms of link-level padding can still be specified
using the primitives in Section 3, by using "leaky pipe" topology to
send the RELAY commands to the Guard node instead of to later nodes in
the circuit.

Because the above link-level padding only sends padding cells if the link is
idle, it can be used in combination with the more complicated circuit-level
padding below, without compounding overhead effects.

3. End-to-end circuit padding

For circuit-level padding, we need the ability to schedule a statistical
distribution of arbitrary padding to overlay on top of non-padding
traffic (aka "Adaptive Padding").

The statistical mechanisms that define padding are known as padding
machines. Padding machines can be hardcoded in Tor, specified in the
consensus, and custom research machines can be listed in Torrc.

3.1. Padding Machines

Circuits can have either one or two state machines at both the origin and at a
specified middle hop.

Each state machine can contain up to three states ("Start", "Burst" and "Gap")
governing their behavior, as well as an "END" state. Not all states need to be

Each state of a padding machine specifies either:
  * A histogram describing inter-arrival cell delays; OR
  * A parameterized delay probability distribution for inter-arrival cell delays

In either case, the lower bound of the delay probability distribution can be
specified as the start_usec parameter, and/or it can be learned by measuring
the RTT of the circuit at the middle node. For client-side machines, RTT
measurement is always set to 0. RTT measurement at the middle node is
calculated by measuring the difference between the time of arrival of an
received cell (ie: away from origin) and the time of arrival of a sent cell
(ie: towards origin). The RTT is continually updated so long as two cells do
not arrive back-to-back in either direction. If the most recent measured RTT
value is larger than our measured value so far, this larger value is used. If
the most recent measured RTT value is lower than our measured value so far, it
is averaged with our current measured value. (We favor longer RTTs slightly in
this way, because circuits are growing away from the middle node and becoming

If the histogram is used, it has an additional special "infinity" bin that
means "infinite delay".

The state can also provide an optional parameterized distribution that
specifies how many total cells (or how many padding cells) can be sent on the
circuit while the machine is in this state, before it transitions to a new

Each state of a padding machine can react to the following cell events:
  * Non-padding cell received
  * Padding cell received
  * Non-padding cell sent
  * Padding cell sent

Additionally, padding machines emit the following internal events to themselves:
  * Infinity bin was selected
  * The histogram bins are empty
  * The length count for this state was exceeded

Each state of the padding machine specifies a set of these events that cause
it to cancel any pending padding, and a set of events that cause it to
transition to another state, or transition back itself.

When an event causes a transition to a state (or back to the same state), a
delay is sampled from the histogram or delay distribution, and padding cell is
scheduled to be sent after that delay.

If a non-padding cell is sent before the timer, the timer is canceled and a
new padding delay is chosen.

3.1.1. Histogram Specification

If a histogram is used by a state (as opposed to a fixed parameterized
distribution), then each of the histograms' fields represent a probability
distribution that is encoded into bins of exponentially increasing width.

The first bin of the histogram (bin 0) has 0 width, with a delay value of
start_usec+rtt_estimate (from the machine definition, and rtt estimate above).

The remaining bins are exponentially spaced, starting at this offset and
covering the range of the histogram, which is range_usec.

The intermediate bins thus divide the timespan range_usec with offset
start_usec+rtt_estimate, so that smaller bin indexes represent narrower time
ranges, doubling up until the last bin. The last bin before the "infinity bin"
thus covers [start_usec+rtt_estimate+range_usec/2,

This exponentially increasing bin width allows the histograms to most
accurately represent small interpacket delay (where accuracy is needed), and
devote less accuracy to larger timescales (where accuracy is not as

To sample the delay time to send a padding packet, perform the
  * Select a bin weighted by the number of tokens in its index compared to
    the total.
  * If the infinity bin is selected, do not schedule padding.
  * If bin 0 is selected, schedule padding at exactly its time value.
  * For other bins, uniformly sample a time value between this bin and
    the next bin, and schedule padding then. Histogram Token Removal

Tokens can be optionally removed from histogram bins whenever a padding or
non-padding packet is sent. With this token removal, the histogram functions
as an overall target delay distribution for the machine while it is in that

If token removal is enabled, when a padding packet is sent, a token is removed
from the bin corresponding to the target delay. When a non-padding packet is
sent, the actual delay from the previous packet is calculated, and the
histogram bin corresponding to that delay is inspected.  If that bin has
tokens remaining, it is decremented.

If the bin has no tokens left, the state removes a token from a different bin,
as specified in its token removal rule. The following token removal options
are defined:
  * None -- Never remove any tokens
  * Exact -- Only remove from the target bin, if it is empty, ignore it.
  * Higher -- Remove from the next higher non-empty bin
  * Lower -- Remove from the next higher non-empty bin
  * Closest -- Remove from the closest non-empty bin by index
  * Closest_time -- Remove from the closest non-empty bin by index, by time

When all bins exept the infinity bin are empty in a histogram, the padding
machine emits the internal "bins empty" event to itself.

Bin 0 and the bin before the infinity bin both have special rules for purposes
of token removal. While removing tokens, all values less than bin 0 are
treated as part of bin 0, and all values greater than
start_usec+rtt_estimate+range_sec are treated as part of the bin before the
infinity bin. Tokens are not removed from the infinity bin when non-padding is
sent. (They are only removed when an "infinite" delay is chosen).

3.1.2. Delay Probability Distribution

Alternatively, a delay probability distribution can be used instead of a
histogram, to sample padding delays.

In this case, the designer also needs to specify the appropriate distribution
parameters, and when a padding cell needs to be scheduled, the padding
subsystem will sample a positive delay value (in microseconds) from that
distribution (where the minimum delay is range_usec+start_usec as is the case
for histograms).

We currently support the following probability distributions:
   Uniform, Logistic, Log-logistic, Geometric, Weibull, Pareto

3.2. State Machine Selection

Clients will select which of the defined available padding machines to use
based on the conditions that these machines specify. These conditions include:
  * How many hops the circuit must be in order for the machine to apply
  * If the machine requires vanguards to be enabled to apply
  * The state the circuit must be in for machines to apply (building,
    relay early cells remaining, opened, streams currently attached).
  * If the circuit purpose matches a set of purposes for the machine.
  * If the target hop of the machine supports circuit padding.

Clients will only select machines whose conditions fully match given circuits.

A machine is represented by a positive number that can be thought of as a "menu
option" through the list of padding machines. The currently supported padding
state machines are:


             A padding machine that obscures the initial circuit setup in an
             attempt to hide onion services.

3.3. Machine Negotiation

When a machine is selected, the client uses leaky-pipe delivery to send a
RELAY_COMMAND_PADDING_NEGOTIATE to the target hop of the machine, using the
following trunnel relay cell payload format:

   * This command tells the relay to alter its min and max netflow
   * timeout range values, and send padding at that rate (resuming
   * if stopped). */
  struct circpad_negotiate {
    u8 version IN [0];

    /** Machine type is left unbounded because we can specify
     * new machines in the consensus */
    u8 machine_type;

Upon receipt of a RELAY_COMMAND_PADDING_NEGOTIATE cell, the middle node sends
a RELAY_COMMAND_PADDING_NEGOTIATED with the following format:

   * This command tells the relay to alter its min and max netflow
   * timeout range values, and send padding at that rate (resuming
   * if stopped). */
  struct circpad_negotiated {
    u8 version IN [0];

    /** Machine type is left unbounded because we can specify
     * new machines in the consensus */
    u8 machine_type;

The 'machine_type' field should be the same as the one from the
PADDING_NEGOTIATE cell. This is because, as an optimization, new machines can
be installed at the client side immediately after tearing down an old machine.
If the response machine type does not match the current machine type, the
response was for a previous machine, and can be ignored.

If the response field is CIRCPAD_RESPONSE_OK, padding was successfully
negotiated. If it is CIRCPAD_RESPONSE_ERR, the machine is torn down and we do
not pad.

4. Examples of Padding Machines

In the original WTF-PAD design[2], the state machines are used as follows:

The "Burst" histogram specifies the delay probabilities for sending a
padding packet after the arrival of a non-padding data packet.

The "Gap" histogram specifies the delay probabilities for sending
another padding packet after a padding packet was just sent from this
node. This self-triggering property of the "Gap" histogram allows the
construction of multi-packet padding trains using a simple statistical

Both "Gap" and "Burst" histograms each have a special "Infinity" bin,
which means "We have decided not to send a packet".

Intuitively, the burst state is used to detect when the line is idle
(and should therefore have few or no tokens in low histogram bins). The
lack of tokens in the low histogram bins causes the system to remain in
the burst state until the actual application traffic either slows,
stalls, or has a gap.

The gap state is used to fill in otherwise idle periods with artificial
payloads from the server (and should have many tokens in low bins, and
possibly some also at higher bins). In this way, the gap state either
generates entirely fake streams of cells, or extends real streams with
additional cells.

The Adaptive Padding Early implementation[3] uses parameterized distributions
instead of histograms, but otherwise uses the states in the same way.

It should be noted that due to our generalization of these states and their
transition possibilities, more complicated interactions are also possible. For
example, it is possible to simulate circuit extension, so that all circuits
appear to continue to extend up until the RELAY_EARLY cell count is depleted.

It is also possible to create machines that simulate traffic on unused
circuits, or mimic onion service activity on clients that aren't otherwise
using onion services.

5. Security considerations and mitigations

The risks from this proposal are primarily DoS/resource exhaustion, and
side channels.

5.1. Rate limiting

Current research[2,3] indicates that padding should be be effective against
website traffic fingerprinting at overhead rates of 50-60%. Circuit setup
behavior can be concealed with far less overhead.

We recommend that three consensus parameters be used in the event that
the network is being overloaded from padding to such a degree that
padding requests should be ignored:

  * circpad_global_max_padding_pct
    - The maximum percent of sent padding traffic out of total traffic
      to allow in a tor process before ceasing to pad. Ex: 75 means
      75 padding packets for every 100 non-padding+padding packets.
      This definition is consistent with the overhead values in Proposal
      #265, though it does not take node position into account.
  * circpad_global_allowed_cells
    - The number of padding cells that must be transmitted before the
      global ratio limit is applied.

Additionally, each machine can specify its own per-machine limits for
the allowed cell counters and padding overhead percentages.

When either global or machine limits are reached, padding is no longer
scheduled. The machine simply becomes idle until the overhead drops below
the threshold.

Finally, the consensus can also be used to specify that clients should
use only machines that are flagged as reduced padding, or disable circuit
padding entirely, with the following two parameters:

   * circpad_padding_reduced=1
     - Tells clients to only use padding machines with the
       'reduced_padding_ok' machine condition flag set.
   * circpad_padding_disabled=1
     - Tells clients to stop circuit padding immediately, and not negotiate
       any further padding machines.

5.2. Overhead accounting

In order to monitor the quantity of padding to decide if we should alter
these limits in the consensus, every node will publish the following
values in a padding-counts line in its extra-info descriptor:

 * read_drop_cell_count
   - The number of RELAY_DROP cells read by this relay.
 * write_drop_cell_count
   - The number of RELAY_DROP cells sent by this relay.

Each of these counters will be rounded to the nearest 10,000 cells. This
rounding parameter will also be listed in the extra-info descriptor line, in
case we change it in a later release.

In the future, we may decide to introduce Laplace Noise in a similar
manner to the hidden service statistics, to further obscure padding

5.3. Side channels

In order to prevent relays from introducing side channels by requesting
padding from clients, all of the padding negotiation commands are only
valid in the outgoing (from the client/OP) direction.

Similarly, to prevent relays from sending malicious padding from arbitrary
circuit positions, if RELAY_DROP cells arrive from a hop other than that
with which padding was negotiated, this cell is counted as invalid for
purposes of CIRC_BW control port fields, allowing the vanguards addon to
close the circuit upon detecting this activity.