Filename: 251-netflow-padding.txt
Title: Padding for netflow record resolution reduction
Authors: Mike Perry
Created: 20 August 2015
Status: Closed
Implemented-In: 0.3.1.1-alpha

NOTE: Please look at section 2 of padding-spec.txt now, not this document.

0. Motivation

 It is common practice by many ISPs to record data about the activity of
 endpoints that use their uplink, if nothing else for billing purposes, but
 sometimes also for monitoring for attacks and general failure.

 Unfortunately, Tor node operators typically have no control over the data
 recorded and retained by their ISP. They are often not even informed about
 their ISP's retention policy, or the associated data sharing policy of those
 records (which tends to be "give them to whoever asks" in practice[1]).

 It is also likely that defenses for this problem will prove useful against
 proposed data retention plans in the EU and elsewhere, since these schemes
 will likely rely on the same technology.

0.1. Background

 At the ISP level, this data typically takes the form of Netflow, jFlow,
 Netstream, or IPFIX flow records. These records are emitted by gateway
 routers in a raw form and then exported (often over plaintext) to a
 "collector" that either records them verbatim, or reduces their granularity
 further[2].

 Netflow records and the associated data collection and retention tools are
 very configurable, and have many modes of operation, especially when
 configured to handle high throughput. However, at ISP scale, per-flow records
 are very likely to be employed, since they are the default, and also provide
 very high resolution in terms of endpoint activity, second only to full packet
 and/or header capture.

 Per-flow records record the endpoint connection 5-tuple, as well as the
 total number of bytes sent and received by that 5-tuple during a particular
 time period. They can store additional fields as well, but it is primarily
 timing and bytecount information that concern us.

 When configured to provide per-flow data, routers emit these raw flow
 records periodically for all active connections passing through them
 based on two parameters: the "active flow timeout" and the "inactive
 flow timeout".

 The "active flow timeout" causes the router to emit a new record
 periodically for every active TCP session that continuously sends data. The
 default active flow timeout for most routers is 30 minutes, meaning that a
 new record is created for every TCP session at least every 30 minutes, no
 matter what. This value can be configured to be from 1 minute to 60 minutes
 on major routers.

 The "inactive flow timeout" is used by routers to create a new record if a
 TCP session is inactive for some number of seconds. It allows routers to
 avoid the need to track a large number of idle connections in memory, and
 instead emit a separate record only when there is activity. This value
 ranges from 10 seconds to 600 seconds on common routers. It appears as
 though no routers support a value lower than 10 seconds.

0.2. Default timeout values of major routers

 For reference, here are default values and ranges (in parenthesis when
 known) for common routers, along with citations to their manuals.

 Some routers speak other collection protocols than Netflow, and in the
 case of Juniper, use different timeouts for these protocols. Where this
 is known to happen, it has been noted.

                         Inactive Timeout              Active Timeout
 Cisco IOS[3]              15s (10-600s)               30min (1-60min)
 Cisco Catalyst[4]         5min                        32min
 Juniper (jFlow)[5]        15s (10-600s)               30min (1-60min)
 Juniper (Netflow)[6,7]    60s (10-600s)               30min (1-30min)
 H3C (Netstream)[8]        60s (60-600s)               30min (1-60min)
 Fortinet[9]               15s                         30min
 MicroTik[10]              15s                         30min
 nProbe[14]                30s                         120s
 Alcatel-Lucent[15]        15s (10-600s)               30min (1-600min)

1. Proposal Overview

 The combination of the active and inactive netflow record timeouts allow us
 to devise a low-cost padding defense that causes what would otherwise be
 split records to "collapse" at the router even before they are exported to
 the collector for storage. So long as a connection transmits data before the
 "inactive flow timeout" expires, then the router will continue to count the
 total bytes on that flow before finally emitting a record at the "active
 flow timeout".

 This means that for a minimal amount of padding that prevents the "inactive
 flow timeout" from expiring, it is possible to reduce the resolution of raw
 per-flow netflow data to the total amount of bytes send and received in a 30
 minute window. This is a vast reduction in resolution for HTTP, IRC, XMPP,
 SSH, and other intermittent interactive traffic, especially when all
 user traffic in that time period is multiplexed over a single connection
 (as it is with Tor).


2. Implementation

 Tor clients currently maintain one TLS connection to their Guard node to
 carry actual application traffic, and make up to 3 additional connections to
 other nodes to retrieve directory information.

 We propose to pad only the client's connection to the Guard node, and not
 any other connection. We propose to treat Bridge node connections to the Tor
 network as client connections, and pad them, but otherwise not pad between
 normal relays.

 Both clients and Guards will maintain a timer for all application (ie:
 non-directory) TLS connections. Every time a non-padding packet is sent or
 received by either end, that endpoint will sample a timeout value from
 between 1.5 seconds and 9.5 seconds. If the connection becomes active for
 any reason before this timer expires, the timer is reset to a new random
 value between 1.5 and 9.5 seconds. If the connection remains inactive until
 the timer expires, a single CELL_PADDING cell will be sent on that connection.

 In this way, the connection will only be padded in the event that it is
 idle, and will always transmit a packet before the minimum 10 second inactive
 timeout.

2.1. Tunable parameters

 We propose that the defense be controlled by the following consensus
 parameters:

   * nf_ito_low
     - The low end of the range to send padding when inactive, in ms.
     - Default: 1500
   * nf_ito_high
     - The high end of the range to send padding, in ms.
     - Default: 9500

   * nf_pad_relays
     - If set to 1, we also pad inactive relay-to-relay connections
     - Default: 0

   * conn_timeout_low
     - The low end of the range to decide when we should close an idle
       connection (not counting padding).
     - Default: 900 seconds after last circuit closes

   * conn_timeout_high
     - The high end of the range to decide when we should close an idle
       connection.
     - Default: 1800 seconds after last circuit close

 If nf_ito_low == nf_ito_high == 0, padding will be disabled.

2.2. Maximum overhead bounds

 With the default parameters, we expect a padded connection to send one
 padding cell every 5.5 seconds (see Appendix A for the statistical analysis
 of expected padding packet rate on an idle link). This averages to 103 bytes
 per second full duplex (~52 bytes/sec in each direction), assuming a 512 byte
 cell and 55 bytes of TLS+TCP+IP headers. For a connection that remains idle
 for a full 30 minutes of inactivity, this is about 92KB of overhead in each
 direction.

 With 2.5M completely idle clients connected simultaneously, 52 bytes per
 second still amounts to only 130MB/second in each direction network-wide,
 which is roughly the current amount of Tor directory traffic[11]. Of course,
 our 2.5M daily users will neither be connected simultaneously, nor entirely
 idle, so we expect the actual overhead to be much lower than this.

2.3. Measuring actual overhead

 To measure the actual padding overhead in practice, we propose to export
 the following statistics in extra-info descriptors for the previous (fixed,
 non-rolling) 24 hour period:

    * Total cells read (padding and non-padding)
    * Total cells written (padding and non-padding)
    * Total CELL_PADDING cells read
    * Total CELL_PADDING cells written
    * Total RELAY_COMMAND_DROP cells read
    * Total RELAY_COMMAND_DROP cells written

 These values will be rounded to 100 cells each, and no values are reported if
 the relay has read or written less than 10000 cells in the previous period.

 RELAY_COMMAND_DROP cells are circuit-level padding not used by this defense,
 but we may as well start recording statistics about them now, too, to aid in
 the development of future defenses.

2.4. Load balancing considerations

 Eventually, we will likely want to update the consensus weights to properly
 load balance the selection of Guard nodes that must carry this overhead.

 We propose that we use the extra-info documents to get a more accurate value
 for the total average Guard and Guard+Exit node overhead of this defense in
 practice, and then use that value to fractionally reduce the consensus
 selection weights for Guard nodes and Guard+Exit nodes, to reflect their
 reduced capacity relative to middle nodes.


3. Threat model and adversarial considerations

 This defense does not assume fully adversarial behavior on the part of the
 upstream network administrator, as that administrator typically has no
 specific interest in trying to deanonymize Tor, but only in monitoring their
 own network for signs of overusage, attack, or failure.

 Therefore, in a manner closer to the "honest but curious" threat model, we
 assume that the netflow collector will be using standard equipment not
 specifically tuned to capturing Tor traffic. We want to reduce the resolution
 of logs that are collected incidentally, so that if they happen to fall into
 the wrong hands, we can be more certain will not be useful.

 We feel that this assumption is a fair one because correlation attacks (and
 statistical attacks in general) will tend to accumulate false positives very
 quickly if the adversary loses resolution at any observation points. It is
 especially unlikely for the the attacker to benefit from only a few
 high-resolution collection points while the remainder of the Tor network
 is only subject to connection-level/per-flow netflow data retention, or even
 less data retention than that.

 Nonetheless, it is still worthwhile to consider what the adversary is capable
 of, especially in light of looming data retention regulation.

 Because no major router appears to have the ability to set the inactive
 flow timeout below 10 seconds, it would seem as though the adversary is left
 with three main options: reduce the active record timeout to the minimum (1
 minute), begin logging full packet and/or header data, or develop a custom
 solution.

 It is an open question to what degree these approaches would help the
 adversary, especially if only some of its observation points implemented
 these changes.

3.1 What about sampled data?

 At scale, it is known that some Internet backbone routers at AS boundaries
 and exchanges perform sampled packet header collection and/or produce
 netflow records based on a subset of the packets that pass through their
 infrastructure.

 The effects of this against Tor were studied before against the (much
 smaller) Tor network as it was in 2007[12]. At sampling rate of 1 out of
 every 2000 packets, the attack did not achieve high accuracy until over
 100MB of data were transmitted, even when correlating only 500 flows in
 a closed-world lab setting.

 We suspect that this type of attack is unlikely to be effective at scale on
 the Tor network today, but we make no claims that this defense will make any
 impact upon sampled correlation, primarily because the amount of padding
 that this defense introduces is comparatively low relative to the amount of
 transmitted traffic that sampled correlation attacks require to attain
 any accuracy.

3.2. What about long-term statistical disclosure?

 This defense similarly does not claim to defeat long-term correlation
 attacks involving many observations over large amounts of time.

 However, we do believe it will significantly increase the amount of traffic
 and the number of independent observations required to attain the same
 accuracy if the adversary uses default per-flow netflow records.

3.3. What about prior information/confirmation?

 In truth, the most dangerous aspect of these netflow logs is not actually
 correlation at all, but confirmation.

 If the adversary has prior information about the location of a target, and/or
 when and how that target is expected to be using Tor, then the effectiveness
 of this defense will be very situation-dependent (on factors such as the
 number of other tor users in the area at that time, etc).

 In any case, the odds that there is other concurrent activity (to
 create a false positive) within a single 30 minute record are much higher
 than the odds that there is concurrent activity that aligns with a
 subset of a series of smaller, more frequent inactive timeout records.


4. Synergistic effects with future padding and other changes

 Because this defense only sends padding when the OR connection is completely
 idle, it should still operate optimally when combined with other forms of
 padding (such as padding for website traffic fingerprinting and hidden service
 circuit fingerprinting). If those future defenses choose to send padding for
 any reason at any layer of Tor, then this defense automatically will not.

 In addition to interoperating optimally with any future padding defenses,
 simple changes to the Tor network usage can serve to further reduce the
 usefulness of any data retention, as well as reduce the overhead from this
 defense.

 For example, if all directory traffic were also tunneled through the main
 Guard node instead of independent directory guards, then the adversary
 would lose additional resolution in terms of the ability to differentiate
 directory traffic from normal usage, especially when it is occurs within
 the same netflow record. As written and specified, the defense will pad
 such tunneled directory traffic optimally.

 Similarly, if bridge guards[13] are implemented such that bridges use their
 own guard node to route all of their connecting client traffic through, then
 users who run bridges will also benefit from blending their own client traffic
 with the concurrent traffic of their connected clients, the sum total of
 which will also be optimally padded such that it only transmits padding when
 the connection to the bridge's guard is completely idle.


Appendix A: Padding Cell Timeout Distribution Statistics

 It turns out that because the padding is bidirectional, and because both
 endpoints are maintaining timers, this creates the situation where the time
 before sending a padding packet in either direction is actually
 min(client_timeout, server_timeout).

 If client_timeout and server_timeout are uniformly sampled, then the
 distribution of min(client_timeout,server_timeout) is no longer uniform, and
 the resulting average timeout (Exp[min(X,X)]) is much lower than the
 midpoint of the timeout range.

 To compensate for this, instead of sampling each endpoint timeout uniformly,
 we instead sample it from max(X,X), where X is uniformly distributed.

 If X is a random variable uniform from 0..R-1 (where R=high-low), then the
 random variable Y = max(X,X) has Prob(Y == i) = (2.0*i + 1)/(R*R).

 Then, when both sides apply timeouts sampled from Y, the resulting
 bidirectional padding packet rate is now a third random variable:
 Z = min(Y,Y).

 The distribution of Z is slightly bell-shaped, but mostly flat around the
 mean. It also turns out that Exp[Z] ~= Exp[X]. Here's a table of average
 values for each random variable:

    R       Exp[X]    Exp[Z]    Exp[min(X,X)]   Exp[Y=max(X,X)]
    2000     999.5    1066        666.2           1332.8
    3000    1499.5    1599.5      999.5           1999.5
    5000    2499.5    2666       1666.2           3332.8
    6000    2999.5    3199.5     1999.5           3999.5
    7000    3499.5    3732.8     2332.8           4666.2
    8000    3999.5    4266.2     2666.2           5332.8
    10000   4999.5    5328       3332.8           6666.2
    15000   7499.5    7995       4999.5           9999.5
    20000   9900.5    10661      6666.2           13332.8

 In this way, we maintain the property that the midpoint of the timeout range
 is the expected mean time before a padding packet is sent in either
 direction.


1. https://lists.torproject.org/pipermail/tor-relays/2015-August/007575.html
2. https://en.wikipedia.org/wiki/NetFlow
3. http://www.cisco.com/en/US/docs/ios/12_3t/netflow/command/reference/nfl_a1gt_ps5207_TSD_Products_Command_Reference_Chapter.html#wp1185203
4. http://www.cisco.com/c/en/us/support/docs/switches/catalyst-6500-series-switches/70974-netflow-catalyst6500.html#opconf
5. https://www.juniper.net/techpubs/software/erx/junose60/swconfig-routing-vol1/html/ip-jflow-stats-config4.html#560916
6. http://www.jnpr.net/techpubs/en_US/junos15.1/topics/reference/configuration-statement/flow-active-timeout-edit-forwarding-options-po.html
7. http://www.jnpr.net/techpubs/en_US/junos15.1/topics/reference/configuration-statement/flow-active-timeout-edit-forwarding-options-po.html
8. http://www.h3c.com/portal/Technical_Support___Documents/Technical_Documents/Switches/H3C_S9500_Series_Switches/Command/Command/H3C_S9500_CM-Release1648%5Bv1.24%5D-System_Volume/200901/624854_1285_0.htm#_Toc217704193
9. http://docs-legacy.fortinet.com/fgt/handbook/cli52_html/FortiOS%205.2%20CLI/config_system.23.046.html
10. http://wiki.mikrotik.com/wiki/Manual:IP/Traffic_Flow
11. https://metrics.torproject.org/dirbytes.html
12. http://freehaven.net/anonbib/cache/murdoch-pet2007.pdf
13. https://gitweb.torproject.org/torspec.git/tree/proposals/188-bridge-guards.txt
14. http://www.ntop.org/wp-content/uploads/2013/03/nProbe_UserGuide.pdf
15. http://infodoc.alcatel-lucent.com/html/0_add-h-f/93-0073-10-01/7750_SR_OS_Router_Configuration_Guide/Cflowd-CLI.html