Introducing Defense against AI-guided Traffic Analysis (DAITA) | Mullvad VPN
mullvad.net
Even if you have encrypted your traffic with a VPN (or the Tor Network), advanced traffic analysis is a growing threat against your privacy. Therefore, we now introduce DAITA.

Even if you have encrypted your traffic with a VPN (or the Tor Network), advanced traffic analysis is a growing threat against your privacy. Therefore, we now introduce DAITA.

Through constant packet sizes, random background traffic and data pattern distortion we are taking the first step in our battle against sophisticated traffic analysis.

  • @[email protected]English
    2•7 months ago

    I think we’re mixing up vocabulary.

    Every IP you talk to is visible to anybody monitoring your network. The sale of net flow data is commonly acknowledged by ISPs. So every computer you talk to is common knowledge for sale.

    In your scenario, let’s say you have five VPN connections set up to go to five endpoints that you control. But if nobody else is using those same endpoints. Your net flow data still exposes exactly what you’re doing. There’s no ambiguity. Your traffic is plainly obvious to anybody observing the network. Even if those VPN connections are adding randomized traffic onto the links.

    • @MigratingtoLemmyEnglish
      1•7 months ago

      Except that I will not necessarily be connecting to the exact same IPs over time, just going to do so in specific ranges which the VPS/colo owns. There’s plenty of people who are going to be renting VPSes and will have their traffic originate from the same IP range as mine, which means that if everybody using TCP had their traffic anonymized like so, the third party wouldn’t actually know that MigratingToLemmy specifically was connecting to AWS at a certain time and from a certain location, so to speak. This hypothesis doesn’t include correlation through other data in the threat model. But it could definitely prevent correlation with traffic across locations, which is similar to what Mullvad states

        • @[email protected]English
          2•6 months ago

          I think you both are talking past each other. You said “But if nobody else is using those same endpoints.” but @[email protected] said “There’s plenty of people who are going to be renting VPSes and will have their traffic originate from the same IP range as mine”. Reading this thread, it seems like you both have different network setups in mind.

          • @[email protected]English
            1•6 months ago

            Thanks for pointing that out. I tried to address that. When I responded about net flow analysis. Having the same IP range as other people does not let you hide in the crowd. The net flow data will identify exact IPs.

            • @[email protected]English
              1•6 months ago

              Hypothetically, what if everybody in the world were using mixnets to obfuscate destination/origin, and then mullvad’s DAITA to obfuscate traffic timing and size. Would netflow analysis be able to defeat that?

              • @[email protected]English
                1•6 months ago

                What is a mix net? Something like TOR? An onion overlay Network where the routing goes between multiple hops before it exits the network?

                Let’s go through a few scenarios first

                Scenario A: you have a link to a common VPN endpoint, that other people use. On this link you generate traffic, a consistent 1 megabyte per second up and down.

                There is now ambiguity about what traffic goes into the VPN, and goes to you. And outside observer would not be able to deduce what traffic is yours just by size and timing.

                This is the gold standard. You remove all possible signal data.

                Scenario B: everyone is using a onion overlay network, and their traffic has a little padding added, and a little extra timing added at every link. This would reduce the probability and outside observer could deduce the entire end to end flow of your traffic. But the type of your traffic could defeat whatever level of obscuring is happening. Imagine you have a real time connection to an network, and you’re typing out Morse code… - - - - sort of thing. Imagine each of those packets has a different size. If I’m observing the network for long enough, I’m going to notice the Morse code type of packets, with the timing and the size go through the onion network. There will be some ambiguity. But enough traffic over enough time would give me high confidence that you’re the source of the traffic. Because the extra obscuring traffic has a probability, but not a guarantee, of masking the shape of your traffic.

                So scenario a is the gold standard, scenario b would be better then nothing. Having a global onion network has its own issues, now you have to trust many nodes instead of one node. All this is down to your threat model and how much effort you’re willing to do.

                • @[email protected]English
                  2•6 months ago

                  Yeah TOR is an example of a mixnet. WHat I was talking about was a combination of your Scenario A and Scenario B, where you have a mixnet where everybody’s traffic goes through multiple proxies, and many people are using each proxy, and you have padding and timing added to make sure traffic flows are consistent. As far as trusting nodes, you have to do that regardless of your set up. If you don’t use any VPN, you have to trust your ISP. If you use a VPN like Mullvad, you have to trust Mullvad. If you use a mixnet, you have to trust that all your chosen proxies aren’t colluding. So like you said, it’s up to your own judgement and threat model.