1) The cost of IPs vs. bandwidth is definitely a function of market offers. Your $500/Gbps/month seems quite expensive compared to what can be found on OVH (which is hosting a large number of relays): they ask ~3 euros/IP/month, including unlimited 100 Mbps traffic. If we assume that wgg = 2/3 and a water level at 10Mbps, this means that, if you want to have 1Gbps of guard bandwidth,
- the current Tor mechanisms would cost you 3 * 10 * 3/2 = 45 euros/month
- the waterfilling mechanism would cost you 3 * 100 = 300 euros/month

The question of what the cheapest attack is can indeed be estimated by looking at market prices for the required resources. Your cost estimate of 3.72 USD/Gbps/month for bandwidth seems off by two orders of magnitude.

The numbers I gave ($2/IP/month and $500/Gbps/month) are the amounts currently charged by my US hosting provider. At the time that I shopped around (which was in 2015), it was by far the best bandwidth cost that I was able to find, and those costs haven’t changed much since then.

Currently on OVH the best I could find for hosting just now was $93.02/per month for 250Mbps unlimited (https://www.ovh.co.uk/dedicated_servers/hosting/1801host01.xml). This yields $372.08/Gbps/month. I am far from certain that this is the best price that one could find - please do point me to better pricing if you have it!

I also just looked at Hetzter - another major Tor-friendly hosting provider. The best I could find was 1Gbps link capped at 100TB/month for $310.49 (https://wiki.hetzner.de/index.php/Traffic/en). 1Gbps sustained upload is 334.8Terabytes (i.e. 1e12 bytes) for a 31-day month. If you exceed that limit, you can arrange to pay $1.24/TB. Therefore I would estimate the cost to be $601.64/Gbps/month. Again, I maybe missing an option more tailored to a high-traffic server, and I would be happy to be pointed to it :-)

Moreover, European bandwidth costs are among the lowest in the world. Other locations are likely to have even higher bandwidth costs (Australia, for example, has notoriously high bandwidth costs).

We do not believe that this is conclusive, as the market changes, and there certainly are dozens of other providers.

I do agree that the market changes, and in fact I expect the cost fo IPs to plummet as the shift to IPv6 becomes pervasive. The current high cost of IPv4 addresses is due to their recent scarcity. In any case, a good question to ask would be how Tor should adjust to changes in market pricing over time.

The same applies for 0-day attacks: if you need to buy them just for attacking Tor, then they are expensive. If you are an organization in the business of handling 0-day attacks for various other reasons, then the costs are very different. And it may be unclear to determine if it is easier/cheaper to compromise 1 top relay or 20 mid-level relays.

I agree that the cost of compromising machines is unclear. However, we should guess, and the business of 0-days has provided some signals of their value in terms of their price. 0-days for the Tor software stack are expensive, as, for security reasons, (well-run) Tor relays run few services other than the tor process. I haven’t seen great data on Linux zero-days, but recently a Windows zero-day (Windows being the second most-common Tor relays OS) appeared to cost $90K (https://www.csoonline.com/article/3077447/security/cost-of-a-windows-zero-day-exploit-this-one-goes-for-90000.html). Deploying a zero-day does impose a cost, as it increases the chance of that exploit being discovered and its value lost. Therefore, such exploits are likely to be deployed only on high-value targets. I would argue that Tor relays are unlikely to be such a target because it is so much cheaper to simply run your own relays. An exception could be a specific targeted investigation in which some suspect is behind a known relay (say, a hidden service behind a guard), because running other relays doesn’t help dislodge the target from behind its existing guard.

And we are not sure that the picture is so clear about botnets either: bots that can become guards need to have high availability (in order to pass the guard stability requirements), and such high availability bots are also likely to have a bandwidth that is higher than the water level (abandoned machines in university networks, ...). As a result, waterfilling would increase the number of high availability bots that are needed, which is likely to be hard.

This doesn’t seem like a good argument to me: “bots that become guards must have high availability, and thus they likely have high bandwidth”. How many bots would become guards in the first place? And why would availability (by which I understand you to mean uptime) imply bandwidth? The economics matter here, and I don’t know too much about botnet economics, but my impressions is that they generally include many thousands of machines and that each bot is generally quickly shut down by its service provider once it starts spewing traffic (i.e. acting as a high-bandwidth Tor relay). Thus waterfilling could benefit botnets by giving them more clients to attack while providing a small amount of bandwidth that falls below the radar of their ISP. This is a speculative argument, I admit, but seems to me to be somewhat more logical than the argument you outlined.

2) Waterfilling makes it necessary for an adversary to run a larger number of relays. Apart from the costs of service providers, this large number of relays need to be managed in an apparently independent way, otherwise they would become suspicious to community members, like nusenu who is doing a great job spotting all anomalies. It seems plausible that running 100 relays in such a way that they look independent is at least as difficult as doing that with 10 relays.

Why is running a large number of relays more noticeable than running a high-bandwidth relay? Actually, it seems, if anything, *less* noticeable. An attacker could even indicate that all the relays are in the same family, and there is no Tor policy that would kick them out of the network for being “too large” of a family. If Tor wants to limit the size of single entities, then they would have to kick out some large existing families (Team Cymru, torservers.net, and the Chaos Communicration Congress come to mind), and moreover such a policy could apply equally well to total amounts of bandwidth as to total number of relays.

3) The question of the protection from relays, ASes or IXPs is puzzling, and we do not have a strong opinion about it. We focused on relays because they are what is available to any attacker, compared to ASes or IXPs which are more specific adversaries. But, if there is a consensus that ASes or IXPs should rather be considered as the main target, it is easy to implement waterfilling at the AS or IXP level rather than at the IP level: just aggregate the bandwidth relayed per AS or IXP, and apply the waterfilling level computation method to them. Or we could mix the weights obtained for all these adversaries, in order to get some improvement against all of them instead of an improvement against only one and being agnostic about the others.

This suggestion of applying waterfilling to individual ASes is intriguing, but would require some a more developed design and argument. Would the attacker model be one that has a fixed cost to compromise/observe a given AS?

4) More fundamentally, since the fundamental idea of Tor is to mix traffic through a large number of relays, it seems to be a sound design principle to make the choice of the critical relays as uniform as possible, as Waterfilling aims to do. A casual Tor user may be concerned to see that his traffic is very likely to be routed through a very small number of top relays, and this effect is likely to increase as soon as a multi-cores compliant implementation of Tor rises (rust dev). Current top relays which suffer from the main CPU bottleneck will probably be free to relay even more bandwidth than they already do, and gain an even more disproportionate consensus weight. Waterfilling might prevent that, and keep those useful relays doing their job at the middle position of paths.