Filtering Trial Results

The Australian Government’s contentious ‘Clean Feed’ internet censorship proposal has got some media attention lately - and by and large, it’s been rightly critical of Senator Conroy’s plans. If you’re not familiar with it, I recommend you read my letters to Conroy and peruse

In the middle of last year, the previous Government commissioned a closed environment testing trial. The results of these were released recently, and the values have been used by both sides to tout the usefulness/uselessness of filters. Handily, these results are available to the public, so I’ve skimmed through the extensive PDF - although I claim no solid understanding of it all - to figure out where the figures are from.

Firstly, a few facts:

  • Six different filtering approaches were tried (with the codenames Alpha, Beta, Gamma, Delta, Theta and Omega).
  • These trials were conducted on a purpose-built network.
  • The network is similar in scale to a Tier 3 ISP.
  • The trials covered speed changes, the effectiveness of blocking blacklisted material, and the valid sites blocked incorrectly.
  • Most filters were only tested against HTTP and HTTPS traffic. Gamma and Omega were also applied to emails, and Delta skipped on HTTPS.

A full grid of numbers is at the bottom of the post, but let’s go through a few comparisons.

Speed vs Blocking

Speed vs Blocking

The speed results here are really mixed. One (Delta) doesn’t drop much at all, but two (Alpha and Gamma) are horrific. All filters manage to block at least 87% of the blacklist - but only Beta comes really close, with 98% (losing a third of the speed in the process though).

Speed vs False Positives

Speed vs False Positives

Note that the scale on the Y Axis drops a bit, but we still get another set of mixed results. None of them are perfect on the false-positives front, and the closest is Gamma on 1.3% - but that comes with severly limited speeds. And really - there are a lot of websites out there. Even 1% covers a fair chunk of the net.

Blocking vs False Positives

Blocking vs False Positives

Here there’s something of a trend, although you have to be looking for it: better blocking effectiveness means a higher number of false positives. That’s not good, people.


There’s really not that much to work off here, no matter what side of the fence you’re on. The main things to keep in mind are:

  • None of the solutions are perfect.
  • All had issues with false-positives
  • This was done on something approaching a Tier 3 ISP - will the performance speeds decrease if we applied these filters on a Tier 1 or 2 ISP? My money’s on yes.
  • It wasn’t Conroy who commissioned this study, so it can’t be pinned against him.
  • Delta, which is arguably the only viable filter judging by performance, still missed 9% of the blacklisted sites.
  • None of the filters were tested against newsgroups, IM (Instant Messaging), or peer-to-peer traffic. I’d imagine HTTP/HTTPS filters are relatively easy, so expecting the same performance and effectiveness for other protocols sounds like a pipe dream to me.

Raw Numbers

*.   *\3. Performance _\2. Effectiveness      
*. Filter *. PPI (Passive Performance Index) *.      
API (Active Performance Index) *. CPI (Change In Performance Index) *.      
BRI (Blocking Rate Index) *. OBI (Overblocking Index)        
Alpha 92% 16% 17% 90% 2.6%
Beta 99% 67% 68% 98% 7.5%
Gamma 98% 14% 14% 87% 1.3%
Delta 99% 98% 100% 91% 2.4%
Theta 78% 76% 99% 95% 7.8%
Omega 101% 79% 78% 94% 2.9%

Glossary of sorts: PPI (Passive Performance Index) is the relative speed when a filter is attached but not running. API (Active Performance Index) is the relative speed when the filter is running. CPI (Change In Performance Index) is API when using PPI as the reference point (instead of uninhibited network speeds). BRI (Blocking Rate Index) is the percentage of blacklisted sites stopped, and OBI (Overblocking Index) is the percentage of friendly sites overzealously blocked.