3 ms·

This method of escrow comes up fairly regularly in the bitcoin world. Each person ignoring the reasons the previous version stopped being promoted. Say I am a

This method of escrow comes up fairly regularly in the bitcoin world. Each person ignoring the reasons the previous version stopped being promoted. Say I am a malicious seller and I am offering a brand new XBox for 1BTC. You the purchaser start the order put up your two BTC, I put up my 1BTC and send you a used XBox only worth 0.75BTC. Now you as the buyer have two choices. You can either a) Refuse to confirm in which case you'll be out 1.25BTC or b) you can accept your loss confirm the order and be out 0.25BTC. Very few people will chose option (a) to the point that it would almost certainly be profitable for the malicious seller. The fall back is reputation systems which require centralization to avoid gaming/having enough data points to be useful.

paulbaumgart 12y ago
Counterevidence: http://en.m.wikipedia.org/wiki/Ultimatum_game http://en.m.wikipedia.org/wiki/Ultimatum_game
- sanswork 12y ago
  Not sure that is actually counter evidence. It says in most cases people will accept unfair splits as long as they aren't too unfair(which is why in my example I send a less valuable replacement instead of no item at all).
  - paulbaumgart 12y ago
    I wonder if it ends up being most suitable for goods where quality is fairly binary, for that reason.
ada1981 12y ago
Even if I put up 2BTC and get nothing, I still have an incentive to approve it to get my 1BTC back. The profitability for the malicious seller is a function of the probability a given buyer in the system will choose to punish the seller or recoup some money. Also, is there a time limit? For example, what if the seller ships and buyer receives but is lazy / not incentived by the deposit.. The challenge here is this is an attempt to create a "Justice Algorithm". I think a reputation based system is probably the solution here, and it doesn't need to be centralized. You could have anonymous reputation nodes -- which would become valueable in the system over time the more trust you have built. Trust seems to be the responsibility of the seller to manufacture by delivering goods as promised. It would be less useful for 1 off transactions with individuals. You'd also want to calculate some trust score based on volume and value and the non-trivial challenge of preventing manufacturing trust via buying / selling to your self.
- mcherm 12y ago
  What might work well would be if the amount put up for escrow were determined in part by the reputation of the entity doing the trade. The entities could still be pseudonymous (it's got an identity, but no one knows who is behind it).
  - sanswork 12y ago
    There is nothing there to stop sock puppet reputation gaming though which is going to be the main attack vector of any distributed anonymous/pseudoanonymous reputation network.
- sanswork 12y ago
  In any distributed trust system you can really only calculate based on volume and value and both of those things are easily manipulated since there is no way to identify individuals. Also you have the situation with your nodes of then requiring an additional layer of trust on the network to determine which nodes are to be trusted. You can't even do things like lower the trust ratings of people who trust people who end up being dishonest because you don't know that they were dishonest in their trust of them. Reputation systems don't work with anonymity. They rarely even work without it.
murbard2 12y ago
Reputation systems do not need centralization to have enough data points, what they need is traction. That can very well be achieved with a decentralized system. But most importantly, the idea of making "trust" graphs using a weak form of transitivity needs to die. Reputation networks should encode sign statements of facts, from which trust can be estimated by participants, using a variety of approaches - in particular non-local approaches which aren't fooled by cliques.
- sanswork 12y ago
  Any ideas for trust approaches that cannot be gamed by cliques in a distributed network? Anything I can think of would break the functionality of the network/remove all usefulness from it or open up other attack vectors.
  - murbard2 12y ago
    You have to differentiate the actual network that holds the data (possibly as a DHT) and the network implicitely described by the statements. The way you avoid clique is by staying within your clique. That is requiring that the entities you extend trust to be trusted by people close to you, who themselves trust each other. But it's not just cliques. Say you have a system like eBay, based on buyer feedback. eBay can be fooled by playing a long game, and selling a lot of cheap items for a while before selling a big item. If instead of eBay you have a distributed store of transaction report, you can constantly refine the trust algorithm. For instance, you might pay attention to the price of the items sold, or to the trustworthiness of the reviewers themselves, etc. The more data goes in, the better your inference. Truth is coherent and entangled. The more data there is, the harder it is to convincingly lie (but the harder it is to remain anonymous as well).
    - sanswork 12y ago
      If you stay within your clique you severely limit the number of possible merchants and if you trust them through personal relationship or foaf why go through the hassle of an escrow system like this? >For instance, you might pay attention to the price of the items sold, or to the trustworthiness of the reviewers themselves, etc. Ok so how about I just sell a bunch of high value "items" between my own accounts? Now my account has participated in a number of valid high value transactions. Since it is distributed there is no way to ensure the transaction is real. >trustworthiness of the reviewers themselves Who build their trustworthiness the same way. This is easy to game look at pagerank for an example of a similar rating system. >The more data there is, the harder it is to convincingly lie This isn't actually true though because you don't verify the data in any way(nor can you) so an attacker can craft the input to tell any story they want.
Animats 12y ago
If you can create new fake identities cheaply, reputation systems can be spammed. Google tried that in 2011, weighting inputs from social networks heavily. They were spammed, badly. A whole ecosystem emerged for spamming social search. Not only was there "Google +1 supply" and "Bulk Likes", there services for fake phone numbers for fake phone number verification appeared. http://www.sitetruth.com/doc/socialisbadforsearch09.pdf http://www.sitetruth.com/doc/socialisbadforsearch09.pdf So, anonymous reputation systems don't work.

paulbaumgart 12y ago

Counterevidence: http://en.m.wikipedia.org/wiki/Ultimatum_game http://en.m.wikipedia.org/wiki/Ultimatum_game

sanswork 12y ago
Not sure that is actually counter evidence. It says in most cases people will accept unfair splits as long as they aren't too unfair(which is why in my example I send a less valuable replacement instead of no item at all).
- paulbaumgart 12y ago
  I wonder if it ends up being most suitable for goods where quality is fairly binary, for that reason.

ada1981 12y ago

Even if I put up 2BTC and get nothing, I still have an incentive to approve it to get my 1BTC back. The profitability for the malicious seller is a function of the probability a given buyer in the system will choose to punish the seller or recoup some money. Also, is there a time limit? For example, what if the seller ships and buyer receives but is lazy / not incentived by the deposit.. The challenge here is this is an attempt to create a "Justice Algorithm". I think a reputation based system is probably the solution here, and it doesn't need to be centralized. You could have anonymous reputation nodes -- which would become valueable in the system over time the more trust you have built. Trust seems to be the responsibility of the seller to manufacture by delivering goods as promised. It would be less useful for 1 off transactions with individuals. You'd also want to calculate some trust score based on volume and value and the non-trivial challenge of preventing manufacturing trust via buying / selling to your self.

mcherm 12y ago
What might work well would be if the amount put up for escrow were determined in part by the reputation of the entity doing the trade. The entities could still be pseudonymous (it's got an identity, but no one knows who is behind it).
- sanswork 12y ago
  There is nothing there to stop sock puppet reputation gaming though which is going to be the main attack vector of any distributed anonymous/pseudoanonymous reputation network.
sanswork 12y ago
In any distributed trust system you can really only calculate based on volume and value and both of those things are easily manipulated since there is no way to identify individuals. Also you have the situation with your nodes of then requiring an additional layer of trust on the network to determine which nodes are to be trusted. You can't even do things like lower the trust ratings of people who trust people who end up being dishonest because you don't know that they were dishonest in their trust of them. Reputation systems don't work with anonymity. They rarely even work without it.

murbard2 12y ago

Reputation systems do not need centralization to have enough data points, what they need is traction. That can very well be achieved with a decentralized system. But most importantly, the idea of making "trust" graphs using a weak form of transitivity needs to die. Reputation networks should encode sign statements of facts, from which trust can be estimated by participants, using a variety of approaches - in particular non-local approaches which aren't fooled by cliques.

sanswork 12y ago
Any ideas for trust approaches that cannot be gamed by cliques in a distributed network? Anything I can think of would break the functionality of the network/remove all usefulness from it or open up other attack vectors.
- murbard2 12y ago
  You have to differentiate the actual network that holds the data (possibly as a DHT) and the network implicitely described by the statements. The way you avoid clique is by staying within your clique. That is requiring that the entities you extend trust to be trusted by people close to you, who themselves trust each other. But it's not just cliques. Say you have a system like eBay, based on buyer feedback. eBay can be fooled by playing a long game, and selling a lot of cheap items for a while before selling a big item. If instead of eBay you have a distributed store of transaction report, you can constantly refine the trust algorithm. For instance, you might pay attention to the price of the items sold, or to the trustworthiness of the reviewers themselves, etc. The more data goes in, the better your inference. Truth is coherent and entangled. The more data there is, the harder it is to convincingly lie (but the harder it is to remain anonymous as well).
  - sanswork 12y ago
    If you stay within your clique you severely limit the number of possible merchants and if you trust them through personal relationship or foaf why go through the hassle of an escrow system like this? >For instance, you might pay attention to the price of the items sold, or to the trustworthiness of the reviewers themselves, etc. Ok so how about I just sell a bunch of high value "items" between my own accounts? Now my account has participated in a number of valid high value transactions. Since it is distributed there is no way to ensure the transaction is real. >trustworthiness of the reviewers themselves Who build their trustworthiness the same way. This is easy to game look at pagerank for an example of a similar rating system. >The more data there is, the harder it is to convincingly lie This isn't actually true though because you don't verify the data in any way(nor can you) so an attacker can craft the input to tell any story they want.

Animats 12y ago

If you can create new fake identities cheaply, reputation systems can be spammed. Google tried that in 2011, weighting inputs from social networks heavily. They were spammed, badly. A whole ecosystem emerged for spamming social search. Not only was there "Google +1 supply" and "Bulk Likes", there services for fake phone numbers for fake phone number verification appeared. http://www.sitetruth.com/doc/socialisbadforsearch09.pdf http://www.sitetruth.com/doc/socialisbadforsearch09.pdf So, anonymous reputation systems don't work.