Fraud proofs


In Bitcoin we have made this tradeoff where we don't have everyone running full nodes. Not everyone is participating equally. If you have a full node, you have lots of gigs of hard drive space, but if you do that, you only get a few transactions per second. Modularizing validation an argument for this is how to improve on this situation.

What's the problem we are trying to solve? This is a real screenshot from one of the Android wallets. What this shows is that the SPV client will believe anything that miners or servers tell them. I don't really own 21 million BTC. All I did to make it think this, was to send it a transaction with 21 million BTC on the outputs. My lite client would have accepted this as valid if it were mined.

Tihs was a trust failure. This phone would have trusted this. Lite clients expect that this error gets caught. When you go back to the original Satoshi whitepaper, when he talked about Simplified Payment Verification, which is instead of checking everything you check only some of the transactions, one strategy would be users accepting alerts from network nodes. A lot of people in the Bitcoin world didn't notice that at the first, but they thought about the idea of warning someone that something is invalid. Maybe hang on, this part of the blockchain is rejected, so move on to the next work chain I see, because that might be the valid one. The most work most valid chain is the chain that is valid, or rather, not most valid but valid is boolean.

How do we know a transaction is valid? I copy and pasted this out of the Bitcoin Core source code. I'll simplify this. Why don't we go and, he talked about a rule that says given that the coin limit is 21 million BTC, surely we should not allow a transaction that creates more than 21 million BTC. That's solid evidence that the transaction is invalid. This rule is enforced. That's why I had to use 2.1 million BTC, not 21 million BTC. "Are you creating more coins than exist?". It can validate that one, but not the next thing, which was to go say, the amount of coins going into the transaction, is greater than the ..... yu can see I just made a mistake there. It's the number of transactions going into the transaction, versus transactions going out of the ... apparently I like inflation in my protocol rules (laughter). If you have this rule that's great, we can go enforce it, what's a valid transaction, obviously I am over-simplifying a heck of a lot. Let's go work with this. How could you make this usable? If I could have this, I could put it into my lite clients and use it. Given that the transaction would be large, maybe I would want a functional style to validate things in parallel. Once you are at that part, what kind of mathematical operation are you checking? Are we making money out of thin air?

When you look at it, the equation you are running is looking like a tree. We are going down the inputs and outputs of a transaction. We are adding them all up, at every point, does the function equal true? For some, is it natural sum? We can store this as, we can do hashes of this data. I think Andrew Miller was the first person to work on this, originally having the idea of merkle sum trees. If I take every step of that computation and hash them into a tree, I have a record of this computation which is in this case one sided addition. If I bring this as a whole transaction, I have my inputs, I have my outputs, I have this rule here at the top, and I can go say, is this transaction valid. What if it isn't? What if somewhere there we create money out of thin air? In this case, I got the order of this right, and apparently money out of thin air is a valid thing. I can identify the problem, since this is all hash, I can now go and delete all the bits that are valid.

So if I go give my phone a copy of this invalid transaction that it knows about that transaction, alright let's walk down, oh hang on that sum didn't work, well, obviously that means the rest of the transaction is invalid, therefore I know to reject any transaction that includes that and claims its valid, like a block maybe. This could be compact, maybe a kilobyte of data. If we structure this in the right way, we can maybe implement the alerts that Satoshi was talking about.

What's neat about this is that when we send data on the wire protocol, we can apply something that looks like compression. You don't need to save the intermediate steps. You just need to give people the data at the bottom. You are not using up any more bandwidth moving the data around. This is something you can use for the clients that supported it.

That sounded good. If we did this, we would have lite clients that wouldn't have to validate anything, they could go to their peers, we could scale Bitcoin up indefinitely. I have done some research into how we would do this. I have gone into a tremendous amount of detail. Some of them are more technical, like how do you commit to a list of committed coins that have been spent or unspent. This commitment structure raises a lot of scalability issues. One of the tradeoffs we would have to make is that given that a transaction is processed atomically, if all inputs are spent in one go, maybe you would split that up and get more parallelism by getting rid of those atomic updates in the protocol. By comparison, that's technical minutatia. Does this actually work?

In a decentralized system, can we pull this off? What if you have missing data? Presumably, we are trying to use this technique to let us go split the blockchain across different nodes. But this is a decentralized network, someone can sybil attack hte network. You don't have a good way of ensuring that someone out there has the data in total. You just happen to see whatever data people are telling you about. If I can create money out of thin air, as a miner why would I give anyone that evidence? Why wouldn't I just leave that data out?

This could apply to transactions, sure. If you look at the transaction, if it's valid but it spends money that was invalid a step back, it's not clear how to force the issue. One approach that might work is a validity challenge. If you remember back when we tried to determine if data was invalid, we can apply the same trick, we can say I suspect the node in this tree is invalid, I don't have data to prove it, please reply with proof, I will assume that stuff I can't see is invalid, therefore I will reject that block.

So you can imagine a system where nodes would collaboratively look for places where the challenges are missing or something. Does this work? If that did work, this is a problem because then I can go to your node, feed oyu validity challenges, and then you would shut down until you have the data again. This doesn't scale well, because you need all the data to accept the block.

I don't know how to solve this. In a centralized environment, if this was a bunch of banks running the blockchain, we can use legal structures to figure out who validates what, maybe Bank of America validates all transactions starting with 1, and Citibank starts validating all transactions starting with 0. But this only works in a centralized environment.

I have done a fair bit of fintech work. I was even employed by R3. Even in that environment, the tech shows some promise, it could be interesting research avenue, does this work in Bitcoin though? It's not really clear.

Another issue is censorship. I can go prove that data is valid. But we have another problem which is that I would like to prove that data is being censored. From the point of view of a miner, they might have honestly not seen a transaction they didn't include. Why didn't they include a transaction? There's no solid evidence of wrongdoing there. They just didn't see your transaction quite likely. If they continued to not mine your transaction, as a human being I could probably come to some kind of agreement that it was a bit suspicious that some donations to Trump are going through or something, but not to Clinton. But I can't write a computer program that does this. It's not a trivial problem. You can see in general that proof-of-work is evidence that things have been published, that something got to an audience.

So maybe you add another proof-of-work layer into this, to insert transactions into the upper blockchain to prove that they should be in shards or lower-levels. But how many layers are we going to add to this? I don't want blockchains within blockchains or whatever. I'm not sure where this is going to go.

This systematic process of looking at these rules in detail, if we can't scale or whatever, we can at least get a good understanding of what the protocol really is. I have also looked at scriptpubkey, this program-based process for deciding transaction validation authorization and signing, and applying this to cryptography in general, like maybe multi-party multi-key thing to sing my emails.

This same approach for scriptpubkey can be applied to validity rules. So at worst we might have a more understandable system, but I don't know if we can do better than that.

Q: Normally, you would have this in, connections so, it's hard to, ...

A: So, the real question is who do they trust? If we are talking about electrum, which will only talk to peers that it knows about. Electrum comes with a list of nodes, like using SSL certs saying this is an electrum server, I know who runs it, and then I connect to 10 others. That's a simple trust model. It's not very decentralized. It's not the dream that Mike Hearn has for making SPV work in a decentralized manner. Bitcoinj on the other hand will connect to people at random, that connecting to 1000 different Bitcoin nodes will probably find a most work Bitcoin blockchain tip. How do you show it's honest? Well when I was doing that attack to show that bitcoinj thought I had sent 2.1 million BTC, well, if I wanted to do this, it would be cheap for me to spin up nodes. Bitcoin Classic is doing this right now to try to show their support. If I have a financial incentive, like making money out of thin air, which I can do from the point of view of someone who aren't running a full node, well I can see that attack working. What about uploading illegal data to make running a full node more dangerous? It's not clear where any of this goes. We'll see as these attacks develop. Right now, under many scenarios, it's very easy.