4 ms·

But how do you get access to large enough sets of training data? Wouldn't that always require plain text access to other peoples mail? Even more, the data would

But how do you get access to large enough sets of training data? Wouldn't that always require plain text access to other peoples mail? Even more, the data would have to be recent, as to take into account trends in approaches.

kuschku 11y ago
Simple: By hosting email for other people. /s But yes, sharing training data (or trained networks) between people is the best solution.
- kbenson 11y ago
  The training data is emails. Sharing training data means your email is no longer contained within your end-to-end encryption, it's leaking allover the place. If you can find a way to extract useful training data from emails while also making it so it doesn't identify anything about you or the emails you've been receiving, I'm sure you can make a lot of money with that. I suspect if it's not impossible, it's extremely hard, and even harder to do right (that is, in a way that we don't find that is later susceptible to some partial reconstruction of attributes).
  - kuschku 11y ago
    Well, there is a way: Share the whole network. Then you combine all networks that others share with you into a new one, and use that yourself. Continuous recombination and cross-breeding of networks is the idea.
    - kbenson 11y ago
      If you share the network, and allow members of the network to use their own data to help train, how do you prevent spammers from joining and submitting garbage, or worse, targeted updates to make specific spam pass?
      - kuschku 11y ago
        Well, the idea is that you can check networks against your own set of organised data — if adding network X reduces the overall effectivity, you just stop using network X and X's score is reduced. EDIT: As HN prevents me from adding new comments right now (Seriously, HN, allow us to post more than 3 comments per hour, it’s seriously hard to hold a conversation like this), I’ll answer your comment here: Users would train networks locally based on their own decisions. Those networks would then be submitted to a repo, and you’d get other networks in return. If a network sorts badly (aka, you always undo its sorts manually), you will not get networks with similar sorting capabilities next time. The concept would automatically prevent people from adding malicious networks – as they’d end up in the local blacklist of users. Obviously you wouldn’t blacklist the network itself, but a representation of its concept of sorting.
        kbenson 11y ago
        So, how are these networks getting their data? Users submitting data? That means users are reducing their individual security to increase the group security as a whole. You are then presented with just consuming this data (and staying secure), or contributing, and we're back at the same point, data needs to be shared so it can be trained against. Let's also look at the incentives for these networks that have data you can subscribe to. How are they supposed to keep spammers out? Any sort of vetting and management of the individual networks will be non-negligible, and if it's not funded will be at a disadvantage to the spammers that are doing this for profit. Finally, I'm not sure that training sets for data like this can be easily combined without a massive amount of reprocessing, if at all. I'm not familiar enough with the classifying networks involved to know, but I suspect that problem alone ranges somewhere from "non-trivial" to "very-hard", if not already solved. It sounds good, and in a perfect world we'd have well run and managed shared networks of fully anonymized spam/phishing classification training data that was easy to combine into individual personal classifiers without having to heavily re-process large training sets. I'm just not sure how feasible the individual parts of that are, much less them combined into a whole.
        thomasahle 11y ago
        I think this idea certainly has interesting parts. We would at least need the following parts: * Neural nets or similar trained models, which we could prove don't leak information on the data they are trained on. * A way to combine these models, without access to their training data, in a way that works as well as training a new model on the union of the data. * A way to exclude spammers. If the models vote on each message, perhaps we would be okay as long as fewer than 50% of the contributors are spammers.
jMyles 11y ago
Presumably you'd have to share training data across instances.

kuschku 11y ago

Simple: By hosting email for other people. /s But yes, sharing training data (or trained networks) between people is the best solution.

kbenson 11y ago
The training data is emails. Sharing training data means your email is no longer contained within your end-to-end encryption, it's leaking allover the place. If you can find a way to extract useful training data from emails while also making it so it doesn't identify anything about you or the emails you've been receiving, I'm sure you can make a lot of money with that. I suspect if it's not impossible, it's extremely hard, and even harder to do right (that is, in a way that we don't find that is later susceptible to some partial reconstruction of attributes).
- kuschku 11y ago
  Well, there is a way: Share the whole network. Then you combine all networks that others share with you into a new one, and use that yourself. Continuous recombination and cross-breeding of networks is the idea.
  - kbenson 11y ago
    If you share the network, and allow members of the network to use their own data to help train, how do you prevent spammers from joining and submitting garbage, or worse, targeted updates to make specific spam pass?
    - kuschku 11y ago
      Well, the idea is that you can check networks against your own set of organised data — if adding network X reduces the overall effectivity, you just stop using network X and X's score is reduced. EDIT: As HN prevents me from adding new comments right now (Seriously, HN, allow us to post more than 3 comments per hour, it’s seriously hard to hold a conversation like this), I’ll answer your comment here: Users would train networks locally based on their own decisions. Those networks would then be submitted to a repo, and you’d get other networks in return. If a network sorts badly (aka, you always undo its sorts manually), you will not get networks with similar sorting capabilities next time. The concept would automatically prevent people from adding malicious networks – as they’d end up in the local blacklist of users. Obviously you wouldn’t blacklist the network itself, but a representation of its concept of sorting.
      - kbenson 11y ago
        So, how are these networks getting their data? Users submitting data? That means users are reducing their individual security to increase the group security as a whole. You are then presented with just consuming this data (and staying secure), or contributing, and we're back at the same point, data needs to be shared so it can be trained against. Let's also look at the incentives for these networks that have data you can subscribe to. How are they supposed to keep spammers out? Any sort of vetting and management of the individual networks will be non-negligible, and if it's not funded will be at a disadvantage to the spammers that are doing this for profit. Finally, I'm not sure that training sets for data like this can be easily combined without a massive amount of reprocessing, if at all. I'm not familiar enough with the classifying networks involved to know, but I suspect that problem alone ranges somewhere from "non-trivial" to "very-hard", if not already solved. It sounds good, and in a perfect world we'd have well run and managed shared networks of fully anonymized spam/phishing classification training data that was easy to combine into individual personal classifiers without having to heavily re-process large training sets. I'm just not sure how feasible the individual parts of that are, much less them combined into a whole.
      - thomasahle 11y ago
        I think this idea certainly has interesting parts. We would at least need the following parts: * Neural nets or similar trained models, which we could prove don't leak information on the data they are trained on. * A way to combine these models, without access to their training data, in a way that works as well as training a new model on the union of the data. * A way to exclude spammers. If the models vote on each message, perhaps we would be okay as long as fewer than 50% of the contributors are spammers.

jMyles 11y ago

Presumably you'd have to share training data across instances.