[Event-level] Noise level and intermediate reporting windows

[alois-bissuel]

Hello,

The noise level is fixed for clicks (ie navigation sources) at 0.24%.

If we follow the randomized response mechanism described in the Differential privacy section, the noise level depends on the number of possible output states of the API, which itself depends on the number of intermediate reporting windows.

If we use an attribution window of less than 2 days, the number of possible state is ((8+3) choose 3) = 165. According to the randomize response mechanism, the noise level with the selected epsilon value of 14 should be 0.01372% (which is far from the 0.24% for three intermediate windows).
For two intermediate reporting windows, the number of possible states is 969 and the noise level should read 0.08051%.

Would it be possible to change the noise level based on the attribution window length? Otherwise I think this puts users of the event-level API who choose it to report only on short attribution windows at a disadvantage.

[csharrison]

This feels very reasonable to me, @alois-bissuel . Thanks for the suggestion, I agree it only makes sense to "pay for" the windows that you could possibly use for a given source.

I think the only major consideration to me is spec/implementation complexity, which we can investigate.

[csharrison]

@alois-bissuel would you be willing to discuss this in the next meeting on Monday?

[alois-bissuel]

With pleasure!

[csharrison]

Great, I think this dovetails nicely into our recent publication of the flexible event-level configurations, which we are hoping to discuss as well:
https://github.com/WICG/attribution-reporting-api/blob/main/flexible_event_config.md

This proposal is a "lite" version of that design which attempts to offer much more flexibility.

[chandan-giri]

Thanks @alois-bissuel for sharing the proposal. This is a very interesting idea to achieve configurability in the event API.

This proposal would be very useful for Google ads as this gives us 2 benefits

1. Reduces Noise: reduce the noise level to 0.08% (for 2 effective windows), 0.01% (for 1 effective window). This is very useful for advertisers with low cvr for which the current noise of 0.24% has a high impact.
2. Configurability: configurability on the last reporting windows to a smaller value than the current global default of 7d. We have seen higher transmission losses for higher values of windows as posted in Updating Reporting Windows for Event-level reports in ARA #736