This video is a very deep dive on all of
the different ways that you can covertly
send and receive data. I'm going to show
you how to secretly store encrypted
files in music that sounds completely
normal [music] to the human ear. I've
even made a little web app where
somebody could paste a link to a Spotify
song and it unlocks a completely [music]
secret separate audio stream. What do
you do
>> with a dead chemist?
>> I'm going to show you how you can send
and receive encrypted messages via
ultrasound with no cell service or Wi-Fi
required by turning your phone into a
modem. We're going to be exploiting
Meta's copyright protection to store
ebooks inside Instagram videos. And I'm
going to show you how I stored, hosted,
and shared malware inside of a harmless
Facebook photo of my friend Gary here.
>> [music]
[music]
>> In 499 BC, [music] Greek ruler Histiaeus
wanted to instigate a revolt against the
Persian king, but sending a traditional
letter was too risky. So, he shaved one
of his servant's heads, tattooed a
message to his scalp, and then waited
for the hair to grow back. He then sent
the servant to enter Persian territory
as an ordinary traveler to visit
Histiaeus' son-in-law where he said,
"Yo, shave my head. Your dad has a
secret message for you." And how exactly
this is more efficient than just having
your servant memorize a secret message
is lost on me, but this may be the first
example of steganography in recorded
history. It also started the Ionian
Revolt and Greco-Persian War, which
lasted over 50 years and killed hundreds
of thousands of people. It turns out
that wax was invented like 500 years
before paper, so people used to jot
stuff down by pouring wax on wood
tablets, then carving in their thoughts,
and rubbing the wax to erase it when
they make a mistake. In the same war
that I just mentioned, Demaratus, an
exiled [music] Greek living in Persia,
heard some gossip of a plot to invade
Greece. So, he scraped off all of the
wax off of a tablet, carved the news
[music] into the wood, and then poured
fresh new wax on the tablet over the
message. It made it through the guards,
and in Sparta, everyone was like, "Why
the did this dude send us a blank
tablet?" until Gorgo, aka the queen of
Sparta, was a little sus about it and
figured out how to read the secret
message. During the Roman Empire, Pliny
the Elder figured out that the
transparent fluid from milkweed plants
charred at much lower temperatures than
wood [music] or paper, making it the
perfect candidate for invisible ink. A
short time after that, Roman military
commanders got sick of picking milkweed
and figured out that their own urine had
the exact same invisible ink properties.
So, for many years, the Romans sent piss
messages back and forth to and from
spies. In 1499, German monk and
mathematician Johannes Trithemius wrote
a series of books called Steganographia,
and they were about magic, basically the
complete idiot's guide to using angels
to communicate over long distances. And
it was actually banned by the Catholic
Church, as I suppose the FCC didn't
allow angel comms at the time. It wasn't
until a century later in 1606 when
people started realizing that some of
these books were actually cover texts
for material on cryptography. These
books were analyzed for give or take 400
years, and it was widely believed that
the third book was actually about angel
communications. And it wasn't until 1996
that third book was finally deciphered
to also be about methods of [music]
hiding information. Post-classical Italy
was full of encryption. Leon Alberti
made the first cipher disk in the late
1400s, [music]
which allowed text to be easily encoded
into random characters and then decoded
by someone else who [music] had that
particular decryption cipher. This is
called a shift cipher or a Caesar
cipher, as this method of encryption was
used by Julius Caesar centuries earlier.
It's pretty simple. If you visualize the
alphabet on a wheel, the number of
shifted keys is the distance between two
[music] sets of letters. I made some
simple HTML code that can encode and
decrypt this easily, [music] and it
would have blown Julius Caesar's mind.
In 1550, mathematician Gerolamo Cardano
published a collection of physical
experiments and inventions called The
Subtlety [music] of Things. And among
these inventions was what is now called
the Cardan grill. It's a simple physical
decryption key that can be placed over
pre-existing text to outline a hidden
message. In 1772, Philip Thicknesse
Philip Thicknesse, what a name. He was a
travel writer, an eccentric, [music] and
a liar, I guess. He wrote a book about
writing music with a harmonic alphabet
as a way to secretly transmit
information. A few decades later,
classical composer Joseph Haydn's little
brother Michael sorted out a simple
musical cipher system where the alphabet
could be represented by notes across a
span of octaves. In 1936, Alan Turing
had an idea for what he called a
universal machine, but it wasn't until
World War II a few years later when
countries started pouring resources into
conceptual computing devices to try and
break this thing, Enigma, the infamous
impossible Nazi encryption puzzle. Fun
fact, while there aren't many fun facts
about Nazis, Enigma was initially
developed by the Dutch to communicate
banking secrets and Germany bought the
patent in 1923. Germany then improved
Enigma and would communicate with troops
using nonsensical codes that thousands
of mathematicians around the world tried
to crack. The machine scrambled letters
using a plugboard, a series of rotors,
and a reflector. Every time a key was
pressed, [music] the machine's internal
wiring shifted, meaning that the same
letter could be encrypted differently
each time. Jerry Don made probably one
of my favorite YouTube videos of all
time demonstrating exactly how Enigma
worked. So, I'm just going to link that
in the description. But military
historians estimate that cracking Enigma
shortened the length of World War II by
nearly half a decade, thus saving
millions of lives and possibly changing
the political map of Europe forever.
Throughout the Cold War, there was a
whole lot of hiding in other
but my favorite were the microdots used
by the KGB and ultimately CIA and
whatever other nefarious organization.
[music]
I actually obtained a KGB replica that a
collector made and it's inside a
quarter. And And you put this quarter
inside of a special wedding ring and
slam it against a desk a few times, it
opens up with a secret container. And
inside that container is this tiny speck
of film that is hardly even visible by
the human eye. And when you view that
film under a powerful microscope, you
have a cipher sheet that was used as a
reference for decrypting communications
from [music] things like Soviet number
stations. And a number station is a
mysterious remote shortwave broadcast
transmitter that repeats recordings
[music] of mysterious codes or text via
Morse code. The radio signal can travel
around the world by bouncing [music] off
of the ionosphere, which acts as a large
mirror for electromagnetic waves. These
signals [music] then can land on an
entirely different continent
communicating the secret message. And
the cipher for that message is printed
on a [music] tiny little micro dot
smaller than a grain of sand hidden
inside of a quarter. And by the way,
creepy number stations still exist
today, although not nearly as much as
they did during the Cold War. In modern
times, encryption standards have gotten
incredibly complex. So complex that I
would have no idea how to understand
them well enough to visualize or explain
them [music] properly in a YouTube
video, but that doesn't matter because
this video isn't about encryption, it's
about hiding in other And my
first experiment is a pretty simple
concept. Could I hide a computer virus
or malware in an image [music] and then
upload that image to Facebook and then
have the malware survive being
recompressed by Facebook's image
handler? Believe it or not, I actually
have a little bit of experience related
to this.
A lot of people get really cranky about
the pronunciation of malware. Like four
out of five people say
>> malware.
>> But I feel weird saying it that way
because I wouldn't say malicious
software. I would say
>> malicious.
>> Anyway.
I created and patented an AI music
detection algorithm that works really
well because it has nothing to do with
detecting AI itself. Let me explain.
When a musician finishes or masters a
song [music] and saves it from Pro Tools
or FL Studio or whatever, the file that
they're uploading to Spotify or YouTube
is uncompressed and lossless. [music]
Meaning every single byte of audio data
is preserved the way that it was saved.
To save on bandwidth, streaming services
like Spotify use a variety of
compression formats to make the files
much, much smaller, [music] but the vast
majority of listeners can't really tell
the difference. This compression method
works using something called discrete
cosine transform, and since AI music
sites typically just ripped music files
from streaming platforms instead of
obtaining the master files from the
artists or paying them for it, my
algorithm essentially forensically
[music]
detects frag ments of that compression
in the training data, and we're going to
use that same math to hide a file in an
image. Discrete cosine transform is used
just about everywhere in your day-to-day
internet life. It's the reason this
video isn't 200 GB, and why 5 minutes
[music] of browsing Instagram doesn't
run you over your cell phone carrier's
data limit. In images, discrete cosine
transform [music] works by deciding what
is useful and what isn't. When you take
a raw photo with an SLR, the camera
measures the electrical charge of each
pixel, [music] estimates the color, and
then stores the data. On this
33-megapixel camera, that's happening
for over 32 million pixels every single
time you take a photo.
I'm going to try and give you a visible
example. Let's take a photo of this
beautiful hen of mine and her new
babies, turn it to grayscale, and then
pull a complicated 8x8 pixel box out of
it. Let's bring it into my little
example program that I whipped up, and
we get to see what the unblended pixels
look like. Over here, we have 64
different coefficients or pixel values,
but we can decrease that all the way
down to 24 before we notice any
meaningful change in [music] the image.
And when you zoom out, those lost
details are going to be completely
unnoticeable to the human eye. This
alone makes the image use 2/3 less data
storage. [music] Now, let's try and
exploit that algorithm to hide a file.
Many forms of steganography in images
utilize a trick called the least [music]
significant bit. When pixel data is
stored in 8-bit images or video, the
color is decided by three ranges of
numbers from 0 to 255,
red, [music] green, and blue. This
allows 16.7 million different
combinations of these three colors,
covering almost all of the colors your
eyes can see in the visible light
spectrum. In fact, for the casual viewer
as demonstrated earlier, this is
actually overkill because you can't tell
the difference between purple with a red
value of 121 or purple with a red value
of 122. Most people don't seem to notice
a change until the color is changed by
four [music] or five values and that
presents a very unique opportunity for
us. So, check this out. If you press a
key on your keyboard, this is how that
character breaks down in binary. To
store that binary into some pixels,
[music] we're going to need to change
the least significant bit, which in this
case will be in red. But, all of these
things that I've talked about in this
section [music]
are at odds with one another because
image and video sites use similar pixel
[music] quantification to reduce file
size, thus corrupting the subtle secret
messages that we could store in the
pixels. There's a really clever and
fascinating workaround for this. JPEG
file compression typically only pays
attention to higher frequencies as our
eyes are less likely to notice the
subtle color or data changes due to the
Weber-Fechner law. I know this is
getting really dorky, but this is cool.
Basically, human brains are really good
at noting subtle differences in smaller
quantities and really shitty at noticing
them in larger quantities. So, if I had
three books on a bookshelf and I added
[music] one, you'd definitely notice
that by glancing at them. If I had 100
books on a bookshelf and added one,
you'd not notice it unless you manually
counted them. The same theory applies
for color [music] frequencies, whether
we're looking at the color itself or
visualizing the waveform. If I were to
increase this frequency by one or two
compressions or rarefactions in the
periodic [music] function, you would
almost certainly notice a difference.
But, if I were to add one or two to this
higher frequency down here, you probably
wouldn't. Fortunately, we have Jabit
Steganography. [music]
It hides the file by targeting those
lower frequencies that JPEG compression
usually ignores and then manually
brightens the entire group of pixels.
The original source code only supports
hiding [music] text or other images, but
a pot of coffee and a few hours later,
I've made a simple web interface that
supports other file [music] formats
including archives. So, let's head over
to Church of Malware and download
nexpose for Microsoft Server Message
Block and run my Javid script with the
web UI. Then, let's take a picture of my
emo dog Gary, drag and drop [music] the
exploit into the image, put in a
password, and adjust some settings, and
upload it to Facebook. Now, [music]
let's download it off of Facebook, pop
it back into Javid, and there you go.
We've hidden and stored a major
dangerous exploit inside of a Facebook
image. By the way, if you just want to
hide stuff locally in your image files
and don't intend on putting it on a
social media service like Facebook where
it'll be recompressed, there's a free
program called OpenStego that does just
this. It even encrypts them. So, for
example, if you wanted to have a text
file with a password or decryption key
in it, you could inject that file into
the Windows desktop [music] shortcut
file for, I don't know, Microsoft Word,
and it could just live inconspicuously
there with only [music] you knowing
about it.
This channel has some history of hiding
data in audio files. I've explored
things like Medicine, I've poison pill
music files to confuse AI models, and
I've even stored data in a bird song.
But, what if, I don't know, you lived in
some type of dictatorship that made
certain types [music] of communication
forbidden? Could we covertly hide data
in audio? The easiest trick to hide data
in audio can be done in any audio
editor. You take a stereo file, and then
you make it mono by making both the left
and right channels identical. Then,
invert the waveform of only one of the
channels, and then paste [music] in an
audio data transmission at a lower
volume. In this case, I'll use Morse
code. When you convert or listen to the
audio file in mono, only the
transmission remains. A more savvy and
complex way of doing this is with
spectral modulation. When you analyze an
audio file, you're likely looking at a
much larger amount of data than your
ears are actually utilizing to listen to
the music or content. And just like with
images, audio compression uses
techniques like discrete [music] cosine
transform to save disk space by removing
what isn't thought to be useful. But,
spectral modulation, on the other hand,
works the opposite way [music] by hiding
random chunks of noise within the audio
spectrum that most people cannot hear. I
wanted to make some software that could
cram enough data into an audio file to
be able to store [music] a completely
separate second secret audio layer.
Before we dive into the different
algorithms and how well they worked,
allow me to introduce you to my pyramid
of expectations. I can definitely hide
stuff in an audio file while preserving
the original music or sounds. I can also
store a decent amount of data in that
audio, and I can do this in a way that
can survive being encoded to more
efficient files and streaming services.
All of these three things are possible,
but one of those will need to be
sacrificed in order to have the other
two. Or think of it like a little slider
that you can just place anywhere in this
triangle. I released a short test
[music] album under a random name with
quite possibly the simplest and most
minimal music possible. It's just me
[music] playing a soft analog
synthesizer melody. The complete lack of
drums or any diversity in the
instrumentation makes [music] hiding
anything extremely difficult. So, this
is that music file ripped from Apple
Music and Tidal with their highest
quality setting enabled.
>> [music]
>> To me, it sounds completely identical to
my original recording, but let's listen
to the hidden stream. What do you do
with a dead chemist?
Bury him. And if you haven't figured it
out, the hidden stream is me telling
science jokes. At higher compression
ratios and the lower quality, like a
360p or 480p YouTube video, it didn't
work quite so well. You can certainly
hear the hidden layer, but you can't
easily make out what I'm saying.
But now I wanted to try hiding an actual
file. And in this case, it would be a
344 kilobyte zipped ebook. To do this,
the encoder converts the audio to a
spectrogram and then applies the exact
same type of technique we used with the
images, [music] hiding data by
manipulating the least significant bit.
This results in this 86-second long
recording being completely
indistinguishable from the original
while being able to store literally any
file under 400 kilobytes. In my tests,
this survived Spotify and Tidal's high
quality settings, but Apple's standard
AAC compression and YouTube either
stripped the data needed to identify the
file or broke the archive. To survive
the lower quality compression ratios on
YouTube videos and streaming platforms,
I added in the option to use
multi-frequency shift keying, which is
much more robust, but can only store
about 2 kilobytes per minute of audio,
and at higher levels could also be more
noticeable in the audio. I included a
stego strength slider so you could test
it out yourself and see what works for
you. If you have or are willing to run
the Python 3 installer, you can easily
launch this yourself on your own web
browser, and of course, if you want to
improve it, all the source code is
there. I have a few more ideas and
programs.
Some years back, Georgi Gerganoff, who
by the way has the best profile photo on
all of GitHub, made a program called GG
Wave, which was a brilliant acoustic
messaging protocol that could run on
Android phones.
>> [music]
>> It hasn't been updated in quite a while
and is formally obsolete, but I figured
his source code was the perfect place to
start. I wanted to update some libraries
so this protocol could run on newer
Android devices, and I wanted to add
some optional ranges: audible,
ultrasound,
and inaudible.
Not all devices will be able to pick up
or transmit the inaudible sounds as it
depends on the frequency response of
your phone's microphone and speakers. I
also added a call sign feature and
encryption. [music] So, for example, if
you were on an airplane with a big group
of friends, you could chat about your
plans for dinner when you land or your
plans [music] to hijack the airplane or
whatever, it's up to you. I added a
fancy little waveform and spectrogram so
you could see the audio signals coming
in, and a loopback test to make [music]
sure that everything is encoding and
decoding correctly on your device. Since
the old Android build of GG Wave was
called Waver, I guess we can call this
one Wavist, which most people will
probably read as Wave Street, which is
probably a cooler name anyway. I did a
lot of testing with Bluetooth [music]
speakers, quietly embedding messages in
media, and ultimately hacking an FM
radio signal and transmitting the
encrypted wavest messages to my car
stereo. This involved setting up a
high-end SDR to receive and decode
[music] FM signals, add in a layer of
ultrasonic messaging noise, and then
encode and transmit it on the same
[music] frequency. By the way, my
transmission wasn't all that powerful,
and I really don't know how legal it is.
I also don't know how useful it is, but
it was fun to see that it worked.
Combining these encrypted communications
[music]
and the steganography earlier in this
video, I ended up just making an
easy-to-use playground called bum
[music] 16, which stands for Ben's
ultrasonic modem that uses 16 FSK or
frequency shift keying by default to
encode [music] and decode messages and
files. It also supports parallel shift
keying for sending data in [music] quiet
environments and forward error
correction for very noisy ones. You can
also select the speed per symbol [music]
to make the data rate slower and more
robust. I've posted this also, it's free
and open source, and there's a rough
build of an Android app that's basically
just [music] a quick conversion of the
web app. So, if you ever want to send
someone a selfie over PA system in a
place with no Wi-Fi or cell service,
>> [music]
>> I guess we've made that possible today.
I found that Meta had published a video
watermarking tool called Video Seal, and
it's used to verify content and its
origin. So, for example, if I upload
something to Instagram, and then
somebody else copies it, Meta will be
able to tell that the content is an
original, and then it will be able to
adjust the algorithm accordingly to
reward the person [music] who created
the content. Pretty cool, seems to work.
It's also used to help identify ethical
AI video problems like [music]
deepfakes. And this mechanism is just
hiding a tiny bit of information in a
video, after all, and it's made for the
sole purpose of surviving social media
file compression. It's also open source
with an MIT license, so this might be
our baby. I borrowed the encoders and
decided to repurpose this into something
that simply cramps data into the raw
output of the Video Seal model. Instead
of embedding codes into random frames of
video, it just utilizes as much of the
video container as possible to maximize
[music] data storage, which ends up
looking [clears throat] like an animated
mixture of static and QR codes. I'm
calling it video zuck, and it's actually
pretty good. I was able to store an ePub
of George Orwell's 1984 and a video file
that was under 2 seconds long, and then
retrieve the data with 100% [music]
accuracy. I uploaded a whole lot of
private Facebook reels to test this out,
and it didn't seem as robust as I had
hoped, but as long as [music] I was
using full HD resolution and large block
sizes, I was repeatedly able to store
and extract an ebook inside of a few
seconds of a Facebook [music] or
Instagram reel. I personally would not
call this a success. After seeing how
much data I could cram into images and
audio, I imagined that with advanced
video encoders, my 4K YouTube videos
could secretly store hundreds of
megabytes. But it turns out, and an
obvious conclusion here, is that
advanced video encoders are advanced at
finding and removing unnecessary or
unwanted data. So if we combined our
most functionally robust audio and video
decoders together, we'd be lucky to get
like 100 kilobytes per second. And the
longer the video, the more error
correction needs to be done, which means
that the storage ratio gets less
efficient as the file gets bigger. Most
importantly and honestly, I wouldn't
call any of these reliable. [music] If
you actually wanted to send someone a
secret message via Spotify or an
Instagram post, you'll need to test
[music] it yourself first to make sure
that it works, as none of these methods
are 100% effective. Now you're
wondering, was this video full of secret
files? There is one in here, and I won't
know if it survived YouTube compression
until it's uploaded. If you learned
something today, enjoyed this video, or
found the code useful, and if you want
to see more stuff like this, or help the
ongoing projects I have with
investigating the efficacy of flock
safety, or building a low-cost device
that can measure and standardize
infrasound noise pollution and air
quality around data centers, or if you
just want to access a bunch of dumb
synthesizers and music and audio assets,
or if you just want access to an amazing
and healthy community of like-minded
folks, my Patreon is for you, and you
can join for as little as $1. Thanks for
watching. Keep creating. Bye.