<s>
Perceptual	B-Application
hashing	I-Application
is	O
the	O
use	O
of	O
a	O
fingerprinting	B-Algorithm
algorithm	I-Algorithm
that	O
produces	O
a	O
snippet	O
,	O
hash	B-Algorithm
,	O
or	O
fingerprint	B-Algorithm
of	O
various	O
forms	O
of	O
multimedia	O
.	O
</s>
<s>
A	O
perceptual	B-Application
hash	I-Application
is	O
a	O
type	O
of	O
locality-sensitive	B-Algorithm
hash	I-Algorithm
,	O
which	O
is	O
analogous	O
if	O
features	B-Algorithm
of	O
the	O
multimedia	O
are	O
similar	O
.	O
</s>
<s>
This	O
is	O
not	O
to	O
be	O
confused	O
with	O
cryptographic	B-Algorithm
hashing	I-Algorithm
,	O
which	O
relies	O
on	O
the	O
avalanche	O
effect	O
of	O
a	O
small	O
change	O
in	O
input	O
value	O
creating	O
a	O
drastic	O
change	O
in	O
output	O
value	O
.	O
</s>
<s>
Perceptual	B-Application
hash	I-Application
functions	O
are	O
widely	O
used	O
in	O
finding	O
cases	O
of	O
online	O
copyright	O
infringement	O
as	O
well	O
as	O
in	O
digital	O
forensics	O
because	O
of	O
the	O
ability	O
to	O
have	O
a	O
correlation	O
between	O
hashes	B-Algorithm
so	O
similar	O
data	O
can	O
be	O
found	O
(	O
for	O
instance	O
with	O
a	O
differing	O
watermark	O
)	O
.	O
</s>
<s>
The	O
1980	O
work	O
of	O
Marr	B-Algorithm
and	I-Algorithm
Hildreth	I-Algorithm
is	O
a	O
seminal	O
paper	O
in	O
this	O
field	O
.	O
</s>
<s>
In	O
June	O
2016	O
Azadeh	O
Amir	O
Asgari	O
published	O
work	O
on	O
robust	O
image	O
hash	B-Algorithm
spoofing	O
.	O
</s>
<s>
Asgari	O
notes	O
that	O
perceptual	B-Application
hash	I-Application
function	O
like	O
any	O
other	O
algorithm	O
is	O
prone	O
to	O
errors	O
.	O
</s>
<s>
Researchers	O
remarked	O
in	O
December	O
2017	O
that	O
Google	B-Application
image	I-Application
search	I-Application
is	O
based	O
on	O
a	O
perceptual	B-Application
hash	I-Application
.	O
</s>
<s>
Research	O
reported	O
in	O
January	O
2019	O
at	O
Northumbria	O
University	O
has	O
shown	O
for	O
video	O
it	O
can	O
be	O
used	O
to	O
simultaneously	O
identify	O
similar	O
contents	O
for	O
video	B-General_Concept
copy	I-General_Concept
detection	I-General_Concept
and	O
detect	O
malicious	O
manipulations	O
for	O
video	O
authentication	O
.	O
</s>
<s>
Research	O
reported	O
in	O
May	O
2020	O
by	O
the	O
University	O
of	O
Houston	O
in	O
deep	O
learning	O
based	O
perceptual	B-Application
hashing	I-Application
for	O
audio	O
has	O
shown	O
better	O
performance	O
than	O
traditional	O
audio	B-Algorithm
fingerprinting	I-Algorithm
methods	O
for	O
the	O
detection	O
of	O
similar/copied	O
audio	O
subject	O
to	O
transformations	O
.	O
</s>
<s>
In	O
addition	O
to	O
its	O
uses	O
in	O
digital	O
forensics	O
,	O
research	O
by	O
a	O
Russian	O
group	O
reported	O
in	O
2019	O
has	O
shown	O
that	O
perceptual	B-Application
hashing	I-Application
can	O
be	O
applied	O
to	O
a	O
wide	O
variety	O
of	O
situations	O
.	O
</s>
<s>
A	O
Chinese	O
team	O
reported	O
in	O
July	O
2019	O
that	O
they	O
had	O
discovered	O
a	O
perceptual	B-Application
hash	I-Application
for	O
speech	O
encryption	O
which	O
proved	O
to	O
be	O
effective	O
.	O
</s>
<s>
A	O
technical	O
summary	O
document	O
,	O
which	O
nicely	O
explains	O
the	O
system	O
with	O
copious	O
diagrams	O
and	O
example	O
photographs	O
,	O
offers	O
that	O
"	O
Instead	O
of	O
scanning	O
images	O
[	O
on	O
corporate ]	O
iCloud	B-Application
[servers],	O
the	O
system	O
performs	O
on-device	O
matching	O
using	O
a	O
database	O
of	O
known	O
CSAM	O
image	O
hashes	B-Algorithm
provided	O
by	O
[	O
the	O
National	O
Center	O
for	O
Missing	O
and	O
Exploited	O
Children ]	O
(	O
NCMEC	O
)	O
and	O
other	O
child-safety	O
organizations	O
.	O
</s>
<s>
Apple	O
further	O
transforms	O
this	O
database	O
into	O
an	O
unreadable	O
set	O
of	O
hashes	B-Algorithm
,	O
which	O
is	O
securely	O
stored	O
on	O
users’	O
devices.	O
"	O
</s>
<s>
In	O
an	O
essay	O
entitled	O
"	O
The	O
Problem	O
With	O
Perceptual	B-Application
Hashes	I-Application
"	O
,	O
Oliver	O
Kuederle	O
produces	O
a	O
startling	O
collision	O
generated	O
by	O
a	O
piece	O
of	O
commercial	O
neural	B-Architecture
net	I-Architecture
software	O
,	O
of	O
the	O
NeuralHash	O
type	O
.	O
</s>
<s>
A	O
photographic	O
portrait	O
of	O
a	O
real	O
woman	O
(	O
Adobe	O
Stock	O
#221271979	O
)	O
reduces	O
through	O
the	O
test	O
algorithm	O
to	O
the	O
same	O
hash	B-Algorithm
as	O
the	O
photograph	O
of	O
a	O
piece	O
of	O
abstract	O
art	O
(	O
from	O
the	O
"	O
deposit	O
photos	O
"	O
database	O
)	O
.	O
</s>
<s>
That	O
is	O
,	O
according	O
to	O
Apple	O
,	O
an	O
Apple	O
employee	O
will	O
then	O
look	O
at	O
your	O
(	O
flagged	O
)	O
pictures	O
...	O
Perceptual	B-Application
hashes	I-Application
are	O
messy	O
.	O
</s>
<s>
Researchers	O
have	O
continued	O
to	O
publish	O
a	O
comprehensive	O
analysis	O
entitled	O
"	O
Learning	O
to	O
Break	O
Deep	O
Perceptual	B-Application
Hashing	I-Application
:	O
The	O
Use	O
Case	O
NeuralHash	O
"	O
,	O
in	O
which	O
they	O
investigate	O
the	O
vulnerability	O
of	O
NeuralHash	O
as	O
a	O
representative	O
of	O
deep	O
perceptual	B-Application
hashing	I-Application
algorithms	O
to	O
various	O
attacks	O
.	O
</s>
<s>
Their	O
results	O
show	O
that	O
hash	B-Algorithm
collisions	O
between	O
different	O
images	O
can	O
be	O
achieved	O
with	O
minor	O
changes	O
applied	O
to	O
the	O
images	O
.	O
</s>
<s>
The	O
authors	O
assume	O
their	O
results	O
to	O
apply	O
to	O
other	O
deep	O
perceptual	B-Application
hashing	I-Application
algorithms	O
as	O
well	O
,	O
questioning	O
their	O
overall	O
effectiveness	O
and	O
functionality	O
in	O
applications	O
such	O
as	O
client-side	O
scanning	O
and	O
chat	O
controls	O
.	O
</s>
