Tor clients and relays make use of CELL_PADDING to reduce the resolution of connection-level metadata retention by ISPs and surveillance infrastructure.
Such metadata retention is implemented by Internet routers in the form of Netflow, jFlow, Netstream, or IPFIX 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.
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 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.
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 15s (10-600s) 30min (1-60min) Cisco Catalyst 5min 32min Juniper (jFlow) 15s (10-600s) 30min (1-60min) Juniper (Netflow)[6,7] 60s (10-600s) 30min (1-30min) H3C (Netstream) 60s (60-600s) 30min (1-60min) Fortinet 15s 30min MicroTik 15s 30min nProbe 30s 120s Alcatel-Lucent 15s (10-600s) 30min (1-600min)
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).
Though flow measurement in principle can be bidirectional (counting cells sent in both directions between a pair of IPs) or unidirectional (counting only cells sent from one IP to another), we assume for safety that all measurement is unidirectional, and so traffic must be sent by both parties in order to prevent record splitting.
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 pad only the client's connection to the Guard node, and not any other connection. We 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 padding packet sent by an endpoint, that endpoint will sample a timeout value from the max(X,X) distribution described in Section 2.3. The default range is from 1.5 seconds to 9.5 seconds time range, subject to consensus parameters as specified in Section 2.6.
(The timing is randomized to avoid making it obvious which cells are padding.)
If another cell is sent for any reason before this timer expires, the timer is reset to a new random value.
If the connection remains inactive until the timer expires, a single CELL_PADDING cell will be sent on that connection (which will also start a new timer).
In this way, the connection will only be padded in a given direction in the event that it is idle in that direction, and will always transmit a packet before the minimum 10 second inactive timeout.
(In practice, an implementation may not be able to determine when,
exactly, a cell is sent on a given channel. For example, even though the
cell has been given to the kernel via a call to
send(2), the kernel may
still be buffering that cell. In cases such as these, implementations
should use a reasonable proxy for the time at which a cell is sent: for
example, when the cell is queued. If this strategy is used,
implementations should try to observe the innermost (closest to the wire)
queue that they practically can, and if this queue is already nonempty,
padding should not be scheduled until after the queue does become empty.)
To limit the amount of padding sent, 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) = (2.0i + 1)/(RR).
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
With the default parameters and the above distribution, we expect a padded connection to send one padding cell every 5.5 seconds. 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 client connection that remains otherwise idle for its expected ~50 minute lifespan (governed by the circuit available timeout plus a small additional connection timeout), this is about 154.5KB of overhead in each direction (309KB total).
With 2.5M completely idle clients connected simultaneously, 52 bytes per second amounts to 130MB/second in each direction network-wide, which is roughly the current amount of Tor directory traffic. 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.
To allow mobile clients to either disable or reduce their padding overhead, the CELL_PADDING_NEGOTIATE cell (tor-spec.txt section 7.2) may be sent from clients to relays. This cell is used to instruct relays to cease sending padding.
If the client has opted to use reduced padding, it continues to send padding cells sampled from the range [9000,14000] milliseconds (subject to consensus parameter alteration as per Section 2.6), still using the Y=max(X,X) distribution. Since the padding is now unidirectional, the expected frequency of padding cells is now governed by the Y distribution above as opposed to Z. For a range of 5000ms, we can see that we expect to send a padding packet every 9000+3332.8 = 12332.8ms. We also half the circuit available timeout from ~50min down to ~25min, which causes the client's OR connections to be closed shortly there after when it is idle, thus reducing overhead.
These two changes cause the padding overhead to go from 309KB per one-time-use Tor connection down to 69KB per one-time-use Tor connection. For continual usage, the maximum overhead goes from 103 bytes/sec down to 46 bytes/sec.
If a client opts to completely disable padding, it sends a CELL_PADDING_NEGOTIATE to instruct the relay not to pad, and then does not send any further padding itself.
Currently, clients negotiate padding only when a channel is created, immediately after sending their NETINFO cell. Recipients SHOULD, however, accept padding negotiation messages at any time.
If a client which previously negotiated reduced, or disabled, padding, and wishes to re-enable default padding (ie padding according to the consensus parameters), it SHOULD send CELL_PADDING_NEGOTIATE START with zero in the ito_low_ms and ito_high_ms fields. (It therefore SHOULD NOT copy the values from its own established consensus into the CELL_PADDING_NEGOTIATE cell.) This avoids the client needing to send updated padding negotiations if the consensus parameters should change. The recipient's clamping of the timing parameters will cause the recipient to use its notion of the consensus parameters.
Clients and bridges MUST reject padding negotiation messages from relays, and close the channel if they receive one.
Connection-level padding is 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 - If nf_ito_low == nf_ito_high == 0, padding will be disabled. * nf_ito_low_reduced - For reduced padding clients: the low end of the range to send padding when inactive, in ms. - Default: 9000 * nf_ito_high_reduced - For reduced padding clients: the high end of the range to send padding, in ms. - Default: 14000 * nf_conntimeout_clients - The number of seconds to keep never-used circuits opened and available for clients to use. Note that the actual client timeout is randomized uniformly from this value to twice this value. - The number of seconds to keep idle (not currently used) canonical channels are open and available. (We do this to ensure a sufficient time duration of padding, which is the ultimate goal.) - This value is also used to determine how long, after a port has been used, we should attempt to keep building predicted circuits for that port. (See path-spec.txt section 2.1.1.) This behavior was originally added to work around implementation limitations, but it serves as a reasonable default regardless of implementation. - For all use cases, reduced padding clients use half the consensus value. - Implementations MAY mark circuits held open past the reduced padding quantity (half the consensus value) as "not to be used for streams", to prevent their use from becoming a distinguisher. - Default: 1800 * nf_pad_before_usage - If set to 1, OR connections are padded before the client uses them for any application traffic. If 0, OR connections are not padded until application data begins. - Default: 1 * nf_pad_relays - If set to 1, we also pad inactive relay-to-relay connections - Default: 0 * nf_conntimeout_relays - The number of seconds that idle relay-to-relay connections are kept open. - Default: 3600