Now let's try and use the trained ONNX file for inference on some manually drawn images on GIMP:
Note that the images must be drawn with white on black. If you use black on white, it the accuracy becomes terrible. This is a good very example of brittleness in AI systems!
By default, the setup runs on CPU only, not GPU, as could be seen by running htop. But by the magic of PyTorch, modifying the program to run on the GPU is trivial:
Since these are so rare, it is still a bit hard to classify them for sure, but they are of great interest no doubt, as as we start to notice these patterns more tend to come if it is a thing.
The CGI comms websites contain the only occurrence of HTTPS, so it might open up the door for a certificate fingerprint as proposed by user joelcollinsdc at: news.ycombinator.com/item?id=36280801!
crt.sh appears to be a good way to look into this:
We've noticed that often when there is a hit range:
there is only one IP for each domain
there is a range of about 20-30 of those
and that this does not seem to be that common. Let's see if that is a reasonable fingerprint or not.
Note that although this is the most common case, we have found multiple hits that viewdns.info maps to the same IP.
First we create a table u (unique) that only have domains which are the only domain for an IP, let's see by how much that lowers the 191 M total unique domains:
time sqlite3 u.sqlite 'create table t (d text, i text)'
time sqlite3 av.sqlite -cmd "attach 'u.sqlite' as u" "insert into u.t select min(d) as d, min(i) as i from t where d not like '%.%.%' group by i having count(distinct d) = 1"
The not like '%.%.%' removes subdomains from the counts so that CGI comms are still included, and distinct in count(distinct is because we have multiple entries at different timestamps for some of the hits.
Let's start with the 208 subset to see how it goes:
time sqlite3 av.sqlite -cmd "attach 'u.sqlite' as u" "insert into u.t select min(d) as d, min(i) as i from t where i glob '208.*' and d not like '%.%.%' and (d like '%.com' or d like '%.net') group by i having count(distinct d) = 1"
OK, after we fixed bugs with the above we are down to 4 million lines with unique domain/IP pairs and which contains all of the original hits! Almost certainly more are to be found!
The numbers of the first column are the IPs as a 32-bit integer representation, which is more useful to search for ranges in.
To make a histogram with the distribution of the single hostname IPs:
#!/usr/bin/env bash
bin=$((2**24))
sqlite3 2013-dns-census-a-novirt.sqlite -cmd '.mode csv' >2013-dns-census-a-novirt-hist.csv <<EOF
select i, sum(cnt) from (
select floor(i/${bin}) as i,
count(*) as cnt
from t
group by 1
union
select *, 0 as cnt from generate_series(0, 255)
)
group by i
EOF
gnuplot \
-e 'set terminal svg size 1200, 800' \
-e 'set output "2013-dns-census-a-novirt-hist.svg"' \
-e 'set datafile separator ","' \
-e 'set tics scale 0' \
-e 'unset key' \
-e 'set xrange[0:255]' \
-e 'set title "Counts of IPs with a single hostname"' \
-e 'set xlabel "IPv4 first byte"' \
-e 'set ylabel "count"' \
-e 'plot "2013-dns-census-a-novirt-hist.csv" using 1:2:1 with labels' \
;
Which gives the following useless noise, there is basically no pattern:
# uniq not amazing as there are often two or three slightly different records repeated on multiple timestamps, but down to 11 GB
python3 mx.py | uniq > mx-uniq.csv
sqlite3 mx.sqlite 'create table t(d text, m text)'
# 13 GB
time sqlite3 mx.sqlite ".import --csv --skip 1 'mx-uniq.csv' t"
# 41 GB
time sqlite3 mx.sqlite 'create index td on t(d)'
time sqlite3 mx.sqlite 'create index tm on t(m)'
time sqlite3 mx.sqlite 'create index tdm on t(d, m)'
# Remove dupes.
# Rows: 150m
time sqlite3 mx.sqlite <<EOF
delete from t
where rowid not in (
select min(rowid)
from t
group by d, m
)
EOF
# 15 GB
time sqlite3 mx.sqlite vacuum
Let's see what the hits use:
awk -F, 'NR>1{ print $2 }' ../media/cia-2010-covert-communication-websites/hits.csv | xargs -I{} sqlite3 mx.sqlite "select distinct * from t where d = '{}'"
At around 267 total hits, only 84 have MX records, and from those that do, almost all of them have exactly:
time sqlite3 mx.sqlite '.mode csv' "attach 'aiddcu.sqlite' as 'av'" '.load ./ip' "select ipi2s(av.t.i), av.t.d from av.t inner join t as mx on av.t.d = mx.d and mx.m = 'mailstore1.secureserver.net' order by av.t.i asc" > avm.csv
where avm stands for av with mx pruning. This leaves us with only ~500k entries left. With one more figerprint we could do a Wayback Machine CDX scanning scan.
Let's check that we still have most our hits in there:
Ciro Santilli 35 Updated 2024-12-23 +Created 1970-01-01
