<s>
Domain	B-Language
generation	I-Language
algorithms	I-Language
(	O
DGA	O
)	O
are	O
algorithms	O
seen	O
in	O
various	O
families	O
of	O
malware	O
that	O
are	O
used	O
to	O
periodically	O
generate	O
a	O
large	O
number	O
of	O
domain	B-Protocol
names	I-Protocol
that	O
can	O
be	O
used	O
as	O
rendezvous	O
points	O
with	O
their	O
command	O
and	O
control	O
servers	O
.	O
</s>
<s>
The	O
large	O
number	O
of	O
potential	O
rendezvous	O
points	O
makes	O
it	O
difficult	O
for	O
law	O
enforcement	O
to	O
effectively	O
shut	O
down	O
botnets	B-Architecture
,	O
since	O
infected	O
computers	O
will	O
attempt	O
to	O
contact	O
some	O
of	O
these	O
domain	B-Protocol
names	I-Protocol
every	O
day	O
to	O
receive	O
updates	O
or	O
commands	O
.	O
</s>
<s>
The	O
use	O
of	O
public-key	B-Application
cryptography	I-Application
in	O
malware	O
code	O
makes	O
it	O
unfeasible	O
for	O
law	O
enforcement	O
and	O
other	O
actors	O
to	O
mimic	O
commands	O
from	O
the	O
malware	O
controllers	O
as	O
some	O
worms	O
will	O
automatically	O
reject	O
any	O
updates	O
not	O
signed	O
by	O
the	O
malware	O
controllers	O
.	O
</s>
<s>
For	O
example	O
,	O
an	O
infected	O
computer	O
could	O
create	O
thousands	O
of	O
domain	B-Protocol
names	I-Protocol
such	O
as	O
:	O
www.xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx1.com	O
and	O
would	O
attempt	O
to	O
contact	O
a	O
portion	O
of	O
these	O
with	O
the	O
purpose	O
of	O
receiving	O
an	O
update	O
or	O
commands	O
.	O
</s>
<s>
The	O
technique	O
was	O
popularized	O
by	O
the	O
family	O
of	O
worms	O
Conficker.a	O
and	O
.b	O
which	O
,	O
at	O
first	O
generated	O
250	O
domain	B-Protocol
names	I-Protocol
per	O
day	O
.	O
</s>
<s>
Starting	O
with	O
Conficker.C	O
,	O
the	O
malware	O
would	O
generate	O
50,000	O
domain	B-Protocol
names	I-Protocol
every	O
day	O
of	O
which	O
it	O
would	O
attempt	O
to	O
contact	O
500	O
,	O
giving	O
an	O
infected	O
machine	O
a	O
1%	O
possibility	O
of	O
being	O
updated	O
every	O
day	O
if	O
the	O
malware	O
controllers	O
registered	O
only	O
one	O
domain	O
per	O
day	O
.	O
</s>
<s>
To	O
prevent	O
infected	O
computers	O
from	O
updating	O
their	O
malware	O
,	O
law	O
enforcement	O
would	O
have	O
needed	O
to	O
pre-register	O
50,000	O
new	O
domain	B-Protocol
names	I-Protocol
every	O
day	O
.	O
</s>
<s>
From	O
the	O
point	O
of	O
view	O
of	O
botnet	B-Architecture
owner	O
,	O
they	O
only	O
have	O
to	O
register	O
one	O
or	O
a	O
few	O
domains	O
out	O
of	O
the	O
several	O
domains	O
that	O
each	O
bot	O
would	O
query	O
every	O
day	O
.	O
</s>
<s>
DGA	O
can	O
also	O
combine	O
words	O
from	O
a	O
dictionary	B-Operating_System
to	O
generate	O
domains	O
.	O
</s>
<s>
These	O
dictionaries	B-Operating_System
can	O
be	O
hard-coded	O
in	O
malware	O
or	O
taken	O
from	O
a	O
publicly	O
accessible	O
source	O
.	O
</s>
<s>
Domains	O
generated	O
by	O
dictionary	B-Operating_System
DGA	O
tend	O
to	O
be	O
more	O
difficult	O
to	O
detect	O
due	O
to	O
their	O
similarity	O
to	O
legitimate	O
domains	O
.	O
</s>
<s>
DGA	O
domain	B-Protocol
names	I-Protocol
can	O
be	O
blocked	O
using	O
blacklists	O
,	O
but	O
the	O
coverage	O
of	O
these	O
blacklists	O
is	O
either	O
poor	O
(	O
public	O
blacklists	O
)	O
or	O
wildly	O
inconsistent	O
(	O
commercial	O
vendor	O
blacklists	O
)	O
.	O
</s>
<s>
Reactionary	O
detection	O
relies	O
on	O
non-supervised	O
clustering	B-Algorithm
techniques	I-Algorithm
and	O
contextual	O
information	O
like	O
network	O
NXDOMAIN	O
responses	O
,	O
WHOIS	B-Protocol
information	O
,	O
and	O
passive	O
DNS	B-Protocol
to	O
make	O
an	O
assessment	O
of	O
domain	O
name	O
legitimacy	O
.	O
</s>
<s>
Recent	O
attempts	O
at	O
detecting	O
DGA	O
domain	B-Protocol
names	I-Protocol
with	O
deep	B-Algorithm
learning	I-Algorithm
techniques	O
have	O
been	O
extremely	O
successful	O
,	O
with	O
F1	B-General_Concept
scores	I-General_Concept
of	O
over	O
99%	O
.	O
</s>
<s>
These	O
deep	B-Algorithm
learning	I-Algorithm
methods	O
typically	O
utilize	O
LSTM	B-Algorithm
and	O
CNN	B-Architecture
architectures	O
,	O
though	O
deep	O
word	B-General_Concept
embeddings	I-General_Concept
have	O
shown	O
great	O
promise	O
for	O
detecting	O
dictionary	B-Operating_System
DGA	O
.	O
</s>
<s>
However	O
,	O
these	O
deep	B-Algorithm
learning	I-Algorithm
approaches	O
can	O
be	O
vulnerable	O
to	O
adversarial	B-General_Concept
techniques	I-General_Concept
.	O
</s>
