I am considering to use DAITA on my phone and was wondering how much the data usage will increase. Can anyone who is using it report on the extra data usage by it? Thanks
It’ll increase by quite a bit. I don’t use DAITA (Defense Against AI-guided Traffic Analysis) but I looked it up a bit more.
As of January 2025, this paper found that mobile devices used about 14MiB of additional bandwidth per website visit. So approximately 7x to 13x more data usage.
However, in March 2025, Mullvad announced that they’re using DAITA V2 which uses about half the amount of dummy packets while maintaining same level of defense.
So considering the earlier study was done during fall of 2024, I’d wager your extra data usage on mobile will be half of what they found, so ~3x to 6x more than if DAITA is turned off.
8 Likes
Very interesting paper, thanks a lot!
Interesting also to see that the overhead with mobile is/was about twice as high as with desktop. I would assume this overhead would also cause a higher battery drain on mobile.
Edit: it seems there was another paper based on v2 here but only in swedish. It is from June and was supervised by Tobias Pulls. Will check if that gives more recent and accurate info.
1 Like
Never wondered about the actual stats, thanks for fetching the numbers, quite insightful! 
I guess it can be a real concern if you’re on a limited mobile plan. 
So here is what I found from the more recent paper (which was done for desktops): DAITA v2 increases the data usage by a factor of 2-3.2 (avg. 2.4), depending on the VPN server. Not sure why the server has an impact though (also no correlation between server location and visited homepage location, all in Sweden, is seen). With v1, the increase was by a factor of 2.8-5.3 (avg. 3.8). Again, depending on the server location. This means that the bandwith overhead was reduced by 62%. Last thing: these values are far lower than the 14 MiB overhead from the first study, as I guess due to the study being done for desktop and not mobile. For v1, an average overhead of 5 MiB is found and for v2 of 1.4 MiB.
However, there are some flaws in the whole comparison: they chose less VPN servers for v2 and more importantly, did not choose the same servers for v1 and v2. V1 showed imho a clear impact of the chosen server on the overhead (see US servers below). Furthermore, the 2 last servers had a huge impact for v1 (factor 4.8 and 5.3) and were not included this time for v2. Excluding those 2 servers from the comparison leads to a bandwidth reduction of 38%. Only 3 servers match between the analysis for v1 and v2 and those show a reduction of 21, -14 and 50% (second one even increased). I plotted this here:
For those interested: the reduction in overhead bandwidth also weakens a bit the protection through DAITA. Resistance to deep fingerprinting reduced by 29 percentage points (50 to 21) and to robust fingerprinting by 15 percentage points (26 to 11).
As far as I have understood: The previous paper mentions that despite the higher data usage by mobile it does give it a weaker protection. But accurate numbers on v2 impact on mobiles seem to be missing..
Still surprised that no one shared here anything from experience - not sure if nobody is using DAITA on mobile or all have unlimited data plans (but then, the battery usage should be noticable I guess?)
EDIT:
Found also a big flaw/uncertainty in the study on mobiles. The sampling time was not identical for mobile and desktop (and as we know, DAITA keeps sending data continuously!):
Furthermore, the differences in performance metrics are also influenced by the capture window used during the data collection phase. For desktop clients, the capture window was either 5 seconds or the time it took for the page to fully load, plus an additional 2 seconds, depending on which occurred first. In contrast, the capture window for mobile clients was always 20 seconds, regardless of how quickly the page loaded.
This discrepancy significantly impacts performance metrics, particularly delay and bandwidth for both sent and received data and potentially DF and RF accuracies.
And from the conclusion:
However, one factor influencing the evaluation is the capture window used during data collection. The longer capture window on mobile devices significantly affected performance metrics and potentially contributed to the higher attack accuracies observed in comparison to desktops.
So.. all in all the findings sound to me quite uncertain and in the end I will likely just try it out and report back.
3 Likes