r12981@catbus: nickm | 2007-05-25 19:23:12 -0400

[tor.git] / doc / incentives.txt

                 Tor Incentives Design Brainstorms

1. Goals: what do we want to achieve with an incentive scheme?

1.1. Encourage users to provide good relay service (throughput, latency).
1.2. Encourage users to allow traffic to exit the Tor network from
     their node.

2. Approaches to learning who should get priority.

2.1. "Hard" or quantitative reputation tracking.

   In this design, we track the number of bytes and throughput in and
   out of nodes we interact with. When a node asks to send or receive
   bytes, we provide service proportional to our current record of the
   node's value. One approach is to let each circuit be either a normal
   circuit or a premium circuit, and nodes can "spend" their value by
   sending and receiving bytes on premium circuits: see section 4.1 for
   details of this design. Another approach (section 4.2) would treat
   all traffic from the node with the same priority class, and so nodes
   that provide resources will get and provide better service on average.

   This approach could be complemented with an anonymous e-cash
   implementation to let people spend reputations gained from one context
   in another context.

2.2. "Soft" or qualitative reputation tracking.

   Rather than accounting for every byte (if I owe you a byte, I don't
   owe it anymore once you've spent it), instead I keep a general opinion
   about each server: my opinion increases when they do good work for me,
   and it decays with time, but it does not decrease as they send traffic.
   Therefore we reward servers who provide value to the system without
   nickle and diming them at each step. We also let them benefit from
   relaying traffic for others without having to "reserve" some of the
   payment for their own use. See section 4.3 for a possible design.

2.3. Centralized opinions from the reputation servers.

   The above approaches are complex and we don't have all the answers
   for them yet. A simpler approach is just to let some central set
   of trusted servers (say, the Tor directory servers) measure whether
   people are contributing to the network, and provide a signal about
   which servers should be rewarded. They can even do the measurements
   via Tor so servers can't easily perform only when they're being
   tested. See section 4.4.

2.4. Reputation servers that aggregate opinions.

   The option above has the directory servers doing all of the
   measurements. This doesn't scale. We can set it up so we have "deputy
   testers" -- trusted other nodes that do performance testing and report
   their results.

   If we want to be really adventurous, we could even
   accept claims from every Tor user and build a complex weighting /
   reputation system to decide which claims are "probably" right.
   One possible way to implement the latter is something similar to
   EigenTrust [http://www.stanford.edu/~sdkamvar/papers/eigentrust.pdf],
   where the opinion of nodes with high reputation more is weighted
   higher.

3. Related issues we need to keep in mind.

3.1. Relay and exit configuration needs to be easy and usable.

   Implicit in all of the above designs is the need to make it easy to
   run a Tor server out of the box. We need to make it stable on all
   common platforms (including XP), it needs to detect its available
   bandwidth and not overreach that, and it needs to help the operator
   through opening up ports on his firewall. Then we need a slick GUI
   that lets people click a button or two rather than editing text files.

   Once we've done all this, we'll hit our first big question: is
   most of the barrier to growth caused by the unusability of the current
   software? If so, are the rest of these incentive schemes superfluous?

3.2. The network effect: how many nodes will you interact with?

   One of the concerns with pairwise reputation systems is that as the
   network gets thousands of servers, the chance that you're going to
   interact with a given server decreases. So if 90% of interactions
   don't have any prior information, the "local" incentive schemes above
   are going to degrade. This doesn't mean they're pointless -- it just
   means we need to be aware that this is a limitation, and plan in the
   background for what step to take next. (It seems that e-cash solutions
   would scale better, though they have issues of their own.)

3.3. Guard nodes

   As of Tor 0.1.1.11, Tor users pick from a small set of semi-permanent
   "guard nodes" for their first hop of each circuit. This seems like it
   would have a big impact on pairwise reputation systems since you
   will only be cashing in on your reputation to a few people, and it is
   unlikely that a given pair of nodes will use each other as guard nodes.

   What does this imply? For one, it means that we don't care at all
   about the opinions of most of the servers out there -- we should
   focus on keeping our guard nodes happy with us.

   One conclusion from that is that our design needs to judge performance
   not just through direct interaction (beginning of the circuit) but
   also through indirect interaction (middle of the circuit). That way
   you can never be sure when your guards are measuring you.

   Both 3.2 and 3.3 may be solved by having a global notion of reputation,
   as in 2.3 and 2.4. However, computing the global reputation from local
   views could be expensive (O(n^2)) when the network is really large.

3.4. Restricted topology: benefits and roadmap.

   As the Tor network continues to grow, we will need to make design
   changes to the network topology so that each node does not need
   to maintain connections to an unbounded number of other nodes. For
   anonymity's sake, we may partition the network such that all
   the nodes have the same belief about the divisions and each node is
   in only one partition. (The alternative is that every user fetches
   his own random subset of the overall node list -- this is bad because
   of intersection attacks.)

   Therefore the "network horizon" for each user will stay bounded,
   which helps against the above issues in 3.2 and 3.3.

   It could be that the core of long-lived servers will all get to know
   each other, and so the critical point that decides whether you get
   good service is whether the core likes you. Or perhaps it will turn
   out to work some other way.

   A special case here is the social network, where the network isn't
   partitioned randomly but instead based on some external properties.
   Social network topologies can provide incentives in other ways, because
   people may be more inclined to help out their friends, and more willing
   to relay traffic if most of the traffic they are relaying comes
   from their friends. It also opens the door for out-of-band incentive
   schemes because of the out-of-band links in the graph.

3.5. Profit-maximizing vs. Altruism.

   There are some interesting game theory questions here.

   First, in a volunteer culture, success is measured in public utility
   or in public esteem. If we add a reward mechanism, there's a risk that
   reward-maximizing behavior will surpass utility- or esteem-maximizing
   behavior.

   Specifically, if most of our servers right now are relaying traffic
   for the good of the community, we may actually *lose* those volunteers
   if we turn the act of relaying traffic into a selfish act.

   I am not too worried about this issue for now, since we're aiming
   for an incentive scheme so effective that it produces tens of
   thousands of new servers.

3.6. What part of the node's performance do you measure?

   We keep referring to having a node measure how well the other nodes
   receive bytes. But don't leeching clients receive bytes just as well
   as servers?

   Further, many transactions in Tor involve fetching lots of
   bytes and not sending very many. So it seems that we want to turn
   things around: we need to measure how quickly a node is _sending_
   us bytes, and then only send it bytes in proportion to that.

   However, a sneaky user could simply connect to a node and send some
   traffic through it, and voila, he has performed for the network. This
   is no good. The first fix is that we only count if you're receiving
   bytes "backwards" in the circuit. Now the sneaky user needs to
   construct a circuit such that his node appears later in the circuit,
   and then send some bytes back quickly.

   Maybe that complexity is sufficient to deter most lazy users. Or
   maybe it's an argument in favor of a more penny-counting reputation
   approach.

   Addendum: I was more thinking of measuring based on who is the service
   provider and service receiver for the circuit. Say Alice builds a
   circuit to Bob. Then Bob is providing service to Alice, since he
   otherwise wouldn't need to spend his bandwidth. So traffic in either
   direction should be charged to Alice. Of course, the same attack would
   work, namely, Bob could cheat by sending bytes back quickly. So someone
   close to the origin needs to detect this and close the circuit, if
   necessary. -JN

3.7. What is the appropriate resource balance for servers vs. clients?

   If we build a good incentive system, we'll still need to tune it
   to provide the right bandwidth allocation -- if we reserve too much
   bandwidth for fast servers, then we're wasting some potential, but
   if we reserve too little, then fewer people will opt to become servers.
   In fact, finding an optimum balance is especially hard because it's
   a moving target: the better our incentive mechanism (and the lower
   the barrier to setup), the more servers there will be. How do we find
   the right balance?

   One answer is that it doesn't have to be perfect: we can err on the
   side of providing extra resources to servers. Then we will achieve our
   desired goal -- when people complain about speed, we can tell them to
   run a server, and they will in fact get better performance.

3.8. Anonymity attack: fast connections probably come from good servers.

   If only fast servers can consistently get good performance in the
   network, they will stand out. "Oh, that connection probably came from
   one of the top ten servers in the network." Intersection attacks over
   time can improve the certainty of the attack.

   I'm not too worried about this. First, in periods of low activity,
   many different people might be getting good performance. This dirties
   the intersection attack. Second, with many of these schemes, we will
   still be uncertain whether the fast node originated the traffic, or
   was the entry node for some other lucky user -- and we already accept
   this level of attack in other cases such as the Murdoch-Danezis attack
   [http://freehaven.net/anonbib/#torta05].

3.9. How do we allocate bandwidth over the course of a second?

   This may be a simple matter of engineering, but it still needs to be
   addressed. Our current token bucket design refills each bucket once a
   second. If we have N tokens in our bucket, and we don't know ahead of
   time how many connections are going to want to send out how many bytes,
   how do we balance providing quick service to the traffic that is
   already here compared to providing service to potential high-importance
   future traffic?

   If we have only two classes of service, here is a simple design:
   At each point, when we are 1/t through the second, the total number
   of non-priority bytes we are willing to send out is N/t. Thus if N
   priority bytes are waiting at the beginning of the second, we drain
   our whole bucket then, and otherwise we provide some delayed service
   to the non-priority bytes.

   Does this design expand to cover the case of three priority classes?
   Ideally we'd give each remote server its own priority number. Or
   hopefully there's an easy design in the literature to point to --
   this is clearly not my field.

   Is our current flow control mechanism (each circuit and each stream
   start out with a certain window, and once they've exhausted it they
   need to receive an ack before they can send more) going to have
   problems with this new design now that we'll be queueing more bytes
   for less preferred nodes? If it turns out we do, the first fix is
   to have the windows start out at zero rather than start out full --
   it will slow down the startup phase but protect us better.

   While we have outgoing cells queued for a given server, we have the
   option of reordering them based on the priority of the previous hop.
   Is this going to turn out to be useful? If we're the exit node (that
   is, there is no previous hop) what priority do those cells get?

   Should we do this prioritizing just for sending out bytes (as I've
   described here) or would it help to do it also for receiving bytes?
   See next section.

3.10. Different-priority cells arriving on the same TCP connection.

   In some of the proposed designs, servers want to give specific circuits
   priority rather than having all circuits from them get the same class
   of service.

   Since Tor uses TCP's flow control for rate limiting, this constraints
   our design choices -- it is easy to give different TCP connections
   different priorities, but it is hard to give different cells on the
   same connection priority, because you have to read them to know what
   priority they're supposed to get.

   There are several possible solutions though. First is that we rely on
   the sender to reorder them so the highest priority cells (circuits) are
   more often first. Second is that if we open two TCP connections -- one
   for the high-priority cells, and one for the low-priority cells. (But
   this prevents us from changing the priority of a circuit because
   we would need to migrate it from one connection to the other.) A
   third approach is to remember which connections have recently sent
   us high-priority cells, and preferentially read from those connections.

   Hopefully we can get away with not solving this section at all. But if
   necessary, we can consult Ed Knightly, a Professor at Rice
   [http://www.ece.rice.edu/~knightly/], for his extensive experience on
   networking QoS.

3.11. Global reputation system: Congestion on high reputation servers?

   If the notion of reputation is global (as in 2.3 or 2.4), circuits that
   go through successive high reputation servers would be the fastest and
   most reliable. This would incentivize everyone, regardless of their own
   reputation, to choose only the highest reputation servers in its
   circuits, causing an over-congestion on those servers.

   One could argue, though, that once those servers are over-congested,
   their bandwidth per circuit drops, which would in turn lower their
   reputation in the future. A question is whether this would overall
   stablize.

   Another possible way is to keep a cap on reputation. In this way, a
   fraction of servers would have the same high reputation, thus balancing
   such load.

3.12. Another anonymity attack: learning from service levels.

   If reputation is local, it may be possible for an evil node to learn
   the identity of the origin through provision of differential service.
   For instance, the evil node provides crappy bandwidth to everyone,
   until it finds a circuit that it wants to trace the origin, then it
   provides good bandwidth. Now, as only those directly or indirectly
   observing this circuit would like the evil node, it can test each node
   by building a circuit via each node to another evil node. If the
   bandwidth is high, it is (somewhat) likely that the node was a part of
   the circuit.

   This problem does not exist if the reputation is global and nodes only
   follow the global reputation, i.e., completely ignore their own view.

3.13. DoS through high priority traffic.

   Assume there is an evil node with high reputation (or high value on
   Alice) and this evil node wants to deny the service to Alice. What it
   needs to do is to send a lot of traffic to Alice. To Alice, all traffic
   from this evil node is of high priority. If the choice of circuits are
   too based toward high priority circuits, Alice would spend most of her
   available bandwidth on this circuit, thus providing poor bandwidth to
   everyone else. Everyone else would start to dislike Alice, making it
   even harder for her to forward other nodes' traffic. This could cause
   Alice to have a low reputation, and the only high bandwidth circuit
   Alice could use would be via the evil node.

3.14. If you run a fast server, can you run your client elsewhere?

   A lot of people want to run a fast server at a colocation facility,
   and then reap the rewards using their cablemodem or DSL Tor client.

   If we use anonymous micropayments, where reputation can literally
   be transferred, this is trivial.

   If we pick a design where servers accrue reputation and can only
   use it themselves, though, the clients can configure the servers as
   their entry nodes and "inherit" their reputation. In this approach
   we would let servers configure a set of IP addresses or keys that get
   "like local" service.

4. Sample designs.

4.1. Two classes of service for circuits.

   Whenever a circuit is built, it is specified by the origin which class,
   either "premium" or "normal", this circuit belongs. A premium circuit
   gets preferred treatment at each node. A node "spends" its value, which
   it earned a priori by providing service, to the next node by sending
   and receiving bytes. Once a node has overspent its values, the circuit
   cannot stay as premium. It can either breaks or converts into a normal
   circuit. Each node also reserves a small portion of bandwidth for
   normal circuits to prevent starvation.

   Pro: Even if a node has no value to spend, it can still use normal
   circuits. This allow casual user to use Tor without forcing them to run
   a server.

   Pro: Nodes have incentive to forward traffic as quick and as much as
   possible to accumulate value.

   Con: There is no proactive method for a node to rebalance its debt. It
   has to wait until there happens to be a circuit in the opposite
   direction.

   Con: A node needs to build circuits in such a way that each node in the
   circuit has to have good values to the next node. This requires
   non-local knowledge and makes circuits less reliable as the values are
   used up in the circuit.

   Con: May discourage nodes to forward traffic in some circuits, as they
   worry about spending more useful values to get less useful values in
   return.

4.2. Treat all the traffic from the node with the same service;
     hard reputation system.

   This design is similar to 4.1, except that instead of having two
   classes of circuits, there is only one. All the circuits are
   prioritized based on the value of the interacting node.

   Pro: It is simpler to design and give priority based on connections,
   not circuits.

   Con: A node only needs to keep a few guard nodes happy to forward their
   traffic.

   Con: Same as in 4.1, may discourage nodes to forward traffic in some
   circuits, as they worry about spending more useful values to get less
   useful values in return.

4.3. Treat all the traffic from the node with the same service;
     soft reputation system.

   Rather than a guaranteed system with accounting (as 4.1 and 4.2),
   we instead try for a best-effort system. All bytes are in the same
   class of service. You keep track of other Tors by key, and give them
   service proportional to the service they have given you. That is, in
   the past when you have tried to push bytes through them, you track the
   number of bytes and the average bandwidth, and use that to weight the
   priority of their connections if they try to push bytes through you.

   Now you're going to get minimum service if you don't ever push bytes
   for other people, and you get increasingly improved service the more
   active you are. We should have memories fade over time (we'll have
   to tune that, which could be quite hard).

   Pro: Sybil attacks are pointless because new identities get lowest
   priority.

   Pro: Smoothly handles periods of both low and high network load. Rather
   than keeping track of the ratio/difference between what he's done for
   you and what you've done for him, simply keep track of what he's done
   for you, and give him priority based on that.

   Based on 3.3 above, it seems we should reward all the nodes in our
   path, not just the first one -- otherwise the node can provide good
   service only to its guards. On the other hand, there might be a
   second-order effect where you want nodes to like you so that *when*
   your guards choose you for a circuit, they'll be able to get good
   performance. This tradeoff needs more simulation/analysis.

   This approach focuses on incenting people to relay traffic, but it
   doesn't do much for incenting them to allow exits. It may help in
   one way through: if there are few exits, then they will attract a
   lot of use, so lots of people will like them, so when they try to
   use the network they will find their first hop to be particularly
   pleasant. After that they're like the rest of the world though. (An
   alternative would be to reward exit nodes with higher values. At the
   extreme, we could even ask the directory servers to suggest the extra
   values, based on the current availability of exit nodes.)

   Pro: this is a pretty easy design to add; and it can be phased in
   incrementally simply by having new nodes behave differently.

4.4. Centralized opinions from the reputation servers.

   Have a set of official measurers who spot-check servers from the
   directory to see if they really do offer roughly the bandwidth
   they advertise. Include these observations in the directory. (For
   simplicity, the directory servers could be the measurers.) Then Tor
   servers give priority to other servers. We'd like to weight the
   priority by advertised bandwidth to encourage people to donate more,
   but it seems hard to distinguish between a slow server and a busy
   server.

   The spot-checking can be done anonymously to prevent selectively
   performing only for the measurers, because hey, we have an anonymity
   network.

   We could also reward exit nodes by giving them better priority, but
   like above this only will affect their first hop. Another problem
   is that it's darn hard to spot-check whether a server allows exits
   to all the pieces of the Internet that it claims to. If necessary,
   perhaps this can be solved by a distributed reporting mechanism,
   where clients that can reach a site from one exit but not another
   anonymously submit that site to the measurers, who verify.

   A last problem is that since directory servers will be doing their
   tests directly (easy to detect) or indirectly (through other Tor
   servers), then we know that we can get away with poor performance for
   people that aren't listed in the directory. Maybe we can turn this
   around and call it a feature though -- another reason to get listed
   in the directory.

5. Recommendations and next steps.

5.1. Simulation.

   For simulation trace, we can use two: one is what we obtained from Tor
   and one from existing web traces.

   We want to simulate all the four cases in 4.1-4. For 4.4, we may want
   to look at two variations: (1) the directory servers check the
   bandwidth themselves through Tor; (2) each node reports their perceived
   values on other nodes, while the directory servers use EigenTrust to
   compute global reputation and broadcast those.

5.2. Deploying into existing Tor network.