Latest / kill switch / how to poison AI music scrapers
Transcript
- 0:11Speaker 1: You mentioned that you weren't really releasing music. Can you
- 0:14Speaker 1: can you tell me about that decision.
- 0:15Speaker 2: I discovered that if you typed in across Suno and
- 0:18Speaker 2: Udio make music like The Flashbulb, it would just sound
- 0:22Speaker 2: like crappy versions of my music.
- 0:25Speaker 1: Ben Jordan is a musician in YouTuber, and he releases
- 0:28Speaker 1: music under the name The flash Bulb. You might not
- 0:31Speaker 1: have heard his music yet, but if you've tried an
- 0:33Speaker 1: AI music generator like Zuno or Judio, you might have
- 0:37Speaker 1: used his music. AI music generators work kind of like
- 0:40Speaker 1: an audio version of chat GBT. You can type in
- 0:43Speaker 1: something like make me a techno song with upbeat vocals
- 0:46Speaker 1: and pianos, and it does it based on a massive
- 0:50Speaker 1: library of music that it's scraped. And when Ben tried
- 0:53Speaker 1: messing around with one of these, he realized that his
- 0:56Speaker 1: music had been scraped into that library too, without his It's.
- 1:01Speaker 2: One of those things where I feel like a lot
- 1:02Speaker 2: of people could type in their name and in might
- 1:05Speaker 2: guess something similar to a song that they made, but like,
- 1:08Speaker 2: this was like undeniable. This was just like, oh, this
- 1:11Speaker 2: is literally just everything that you would expect in a
- 1:15Speaker 2: song of mine, all the even the weird things except
- 1:18Speaker 2: for the me part.
- 1:21Speaker 1: So Bin came up with a solution, a program that
- 1:24Speaker 1: adds imperceptible noise to a music track, confusing AI models
- 1:29Speaker 1: and preventing them from replicating the track. This is a
- 1:31Speaker 1: technique called poison pilling.
- 1:33Speaker 2: Poison pilling started with images. There's one called night shade,
- 1:38Speaker 2: and what it did is it essentially just generated some
- 1:41Speaker 2: stuff in the images that was mostly invisible to humans,
- 1:45Speaker 2: and then the AI would see it as something else,
- 1:47Speaker 2: or it would confuse it.
- 1:49Speaker 1: A couple of years ago, as l limbs were really
- 1:51Speaker 1: taken off, a group of researchers at the University of
- 1:54Speaker 1: Chicago developed night Shade and glaze. These are programs that
- 1:58Speaker 1: take an image and make tiny changes to them. These
- 2:01Speaker 1: changes are basically imperceptible to the human eye, but it
- 2:05Speaker 1: would confuse an AI model. The thinking was that if
- 2:08Speaker 1: artists applied nightshade to their images and those images were
- 2:12Speaker 1: then scraped to train AI models, it would not only
- 2:15Speaker 1: prevent the l elms from learning anything from the individual
- 2:18Speaker 1: artists work, but it would also quote poison the data
- 2:21Speaker 1: sets and make those models less reliable.
- 2:24Speaker 2: And so I was like, Okay, well, how possible is
- 2:27Speaker 2: this it's music because I know that adversarial noise attacks
- 2:30Speaker 2: are possible on things like your Google Home or Oxa
- 2:34Speaker 2: or Siri, and it turns out that it is totally possible.
- 2:43Speaker 1: I'm afraid from Kaleidoscope and iHeart podcasts. This is kill switch.
- 2:54Speaker 1: I'm defter, Thomas. I'm sorry. When was the first time
- 3:28Speaker 1: that you really started feeling AI impacting your music personally?
- 3:34Speaker 2: The beginning would be almost in a positive exploratory way,
- 3:38Speaker 2: in like twenty sixteen, Google release Magenta, and in one
- 3:42Speaker 2: of my albums, I used it to sort of generate
- 3:45Speaker 2: this weird morphing sound between three different instruments, and it
- 3:49Speaker 2: was like unlike anything I had really heard before, and
- 3:51Speaker 2: it was this new type synthesis. So of course I
- 3:52Speaker 2: like jumped all over it, and I I was fascinated
- 3:54Speaker 2: with it, and then you know, moving on from there.
- 3:58Speaker 2: It's just the current landscape that we're in that makes
- 4:00Speaker 2: it so bad. So for example, with Spotify refusing to
- 4:04Speaker 2: pay less than a thousand streams per song, you have
- 4:07Speaker 2: like a year to get a thousand streams. If you
- 4:08Speaker 2: don't get that, they don't pay you. So you have
- 4:10Speaker 2: all this already, like low royalties being paid out by
- 4:14Speaker 2: these digital streaming platforms, and you have way too many
- 4:17Speaker 2: artists already, you know, for the system to actually work
- 4:20Speaker 2: where people would be making a living. And now you
- 4:23Speaker 2: have people who aren't musicians who are just using these
- 4:26Speaker 2: services to generate as many songs as they possibly can
- 4:29Speaker 2: and in their monthly subscription.
- 4:32Speaker 1: AI generated music is starting to creep into music platforms
- 4:36Speaker 1: like Spotify and even YouTube. Now you might have come
- 4:39Speaker 1: across it at this point, and maybe you recognized it,
- 4:42Speaker 1: and maybe you didn't. Aside from being really annoying for
- 4:45Speaker 1: people who actually care about music, for musicians, this is
- 4:49Speaker 1: taking away attention and thus money, because a lot of artists'
- 4:53Speaker 1: income depends on the number of streams they get. And
- 4:56Speaker 1: it also doesn't help that the CEO of Suno, which
- 4:59Speaker 1: is one of the most popular AI music companies right now,
- 5:02Speaker 1: doesn't seem to really appreciate the music creation process.
- 5:06Speaker 3: It's not really enjoyable to make music now people enjoy it.
- 5:11Speaker 3: It takes a lot of time, it takes a lot
- 5:12Speaker 3: of practice. You need to get really good at an
- 5:14Speaker 3: instrument or really good at a piece of production software.
- 5:17Speaker 3: I think the majority of people don't enjoy the majority
- 5:19Speaker 3: of the time they spend making music.
- 5:23Speaker 2: That's like one of the most absurd things I've ever
- 5:25Speaker 2: heard in my life. I mean, that's how I hear
- 5:27Speaker 2: that is a CEO trying to justify the existence of
- 5:31Speaker 2: a company in a practical business sense.
- 5:34Speaker 1: So before he went the poison pill route, Bin's first
- 5:37Speaker 1: idea was to make something that would detect if music
- 5:40Speaker 1: was generated by AI. That way, a platform like Spotify
- 5:43Speaker 1: could use it to just reject any AI generated music
- 5:46Speaker 1: that someone tried to upload.
- 5:48Speaker 2: Basically, when you put something on Spotify or really anywhere
- 5:52Speaker 2: where you listen to music generally, obviously, file size and
- 5:56Speaker 2: bandwidth are giant considerations, and so you have to compress
- 5:59Speaker 2: all and within that you use techniques like inverted discrete
- 6:05Speaker 2: cosine transform and a bunch of smart sounding things, and
- 6:08Speaker 2: you can detect discrete cosine transform. And so the thing
- 6:13Speaker 2: is is that Suno and Udio, they allegedly went on
- 6:16Speaker 2: YouTube and Spotify and they just scraped and scraped and
- 6:19Speaker 2: scraped and learned and learned and learned, and so it's
- 6:22Speaker 2: quite easy to detect it.
- 6:23Speaker 1: I see what you're saying. So you're you were able
- 6:25Speaker 1: to basically detect Okay, this was downloaded from Spotify because
- 6:30Speaker 1: it's basically a sound signature. Even if a person doesn't
- 6:34Speaker 1: hear it, computer wise, you can sell YEP.
- 6:36Speaker 2: And so the sort of idea after I announced that,
- 6:40Speaker 2: I got a lot of people saying, yeah, but now
- 6:41Speaker 2: they're just going to use raw wave files, they're just
- 6:43Speaker 2: going to use the masters, and it's like good. Then
- 6:45Speaker 2: they have to negotiate with the artist like that opens
- 6:48Speaker 2: a conversation.
- 6:49Speaker 1: The masters we're talking about here are the original, highest
- 6:53Speaker 1: quality track, which would be owned by the artist or
- 6:55Speaker 1: the label. But in order to use those master files,
- 6:58Speaker 1: you need to get them directly from the person who
- 7:01Speaker 1: owns them, and that would mean you'd probably need to
- 7:04Speaker 1: pay them. The goal isn't necessarily to stop AI, just
- 7:08Speaker 1: to stop AI that the artist isn't getting paid for.
- 7:11Speaker 1: And Ben thought this project could dissuade AI companies from
- 7:14Speaker 1: scraping data because platforms could use this tool to detect
- 7:18Speaker 1: and again then reject that AI music. But Spotify hasn't
- 7:22Speaker 1: implemented this, and as far as we know, it hasn't
- 7:25Speaker 1: stopped AI companies from continuing to scrape music.
- 7:29Speaker 2: I mean, tell Sam Altman to pay for all the
- 7:33Speaker 2: data that he's training everything with them. So, so I
- 7:35Speaker 2: guess that's what led to the next step, right, It's like, okay, well,
- 7:38Speaker 2: how do we prevent it from being trained?
- 7:41Speaker 1: Enter the poison pill. Ben knew that you could do
- 7:45Speaker 1: this in images, but how would this work in audio?
- 7:48Speaker 1: It turns out the process is actually pretty similar.
- 7:52Speaker 2: It's actually not all that different from a technical standpoint,
- 7:55Speaker 2: because the majority of AI music sites, how they're really
- 7:59Speaker 2: working is it's all based on the original unit model
- 8:02Speaker 2: that was for microscopic imaging. That's sort of like what
- 8:05Speaker 2: changed this whole generative AI thing and made it so
- 8:09Speaker 2: much easier to train things than it used to be.
- 8:10Speaker 2: But if you've ever seen an audio spectrogram that you
- 8:13Speaker 2: would see like a violin, it would be like yeah,
- 8:16Speaker 2: and so you would see the note or a line
- 8:19Speaker 2: slowly get thicker and thicker as the violin got louder,
- 8:22Speaker 2: where as a guitar or a piano it would be
- 8:26Speaker 2: an instant star snare drum would be like an instant
- 8:29Speaker 2: start and then maybe a little bit of fade out
- 8:31Speaker 2: on the end depending on how it's mixed. You know,
- 8:34Speaker 2: there are programs, for example, where you could listen to
- 8:37Speaker 2: audio that you draw, and so it's basically doing that.
- 8:39Speaker 2: It's just reading the spectrogram of the audio and then
- 8:43Speaker 2: learning from that and then re encoding another spectrum and
- 8:46Speaker 2: then converting that back into audio. So it's kind of
- 8:49Speaker 2: funny because that's like that's a little bit of a
- 8:51Speaker 2: flawed way of generating audio to begin with. Like you
- 8:56Speaker 2: could usually hear something if it's been converted to a
- 8:58Speaker 2: spectrum in back, and you can hear that almost all
- 9:00Speaker 2: EI music. That's kind of why it sounds a little
- 9:02Speaker 2: glitchy or squeaky or not. It's hard to describe.
- 9:06Speaker 1: I didn't realize that. The way that lms are essentially,
- 9:11Speaker 1: you know these models are interpreting music, is they're really
- 9:13Speaker 1: interpreting them as images. Yeah, using this technique, been created
- 9:19Speaker 1: a process he calls Poisonify. When you run a track
- 9:22Speaker 1: through Poisonify, it'll add noise. It's imperceptible to us but
- 9:27Speaker 1: visible in the audio spectrogram. This confuses the AI training
- 9:31Speaker 1: on it to the point where it can't identify instruments.
- 9:35Speaker 2: So Poisonify is essentially preventing what Magenta initially did, where
- 9:40Speaker 2: it learns primarily can identify instruments and can identify style
- 9:45Speaker 2: and things like that. It cloaks it to where it
- 9:47Speaker 2: just thinks that it's tearing something else. You could have
- 9:49Speaker 2: targeted attacks where you can say I want my piano
- 9:54Speaker 2: to sound like a harmonica or something, or you could
- 9:58Speaker 2: have untargeted attacks, or it'll just kind of go with
- 10:00Speaker 2: whatever's easiest. And when using them in a particolar way,
- 10:04Speaker 2: you can successfully make these instruments and these styles unidentifiable.
- 10:09Speaker 2: So Sono or Udio they would get confused. Then when
- 10:12Speaker 2: it came time to draw a new spectrogram to convert
- 10:16Speaker 2: into audio, they would probably draw some of the wrong ones.
- 10:20Speaker 1: So this means that you could put say an e
- 10:22Speaker 1: EDM track that's treated with Poisonify into Suno, for example,
- 10:26Speaker 1: and if you ask Suno to generate something similar or
- 10:29Speaker 1: extend the song, it'll spit out something totally unrelated, like
- 10:33Speaker 1: acoustic guitar music. Here's a clip from Ben testing that
- 10:36Speaker 1: out with his own music in his YouTube video about
- 10:39Speaker 1: this process. So here we go.
- 10:41Speaker 2: We can upload my original song here and now here
- 10:49Speaker 2: is SUNO's AI extension of that song, Sweedish.
- 11:00Speaker 1: Glowing.
- 11:03Speaker 2: Okay, now let's upload my poison if encoded track and
- 11:13Speaker 2: here is SUNO's AI generated extension. I would describe this
- 11:22Speaker 2: as music from an Airport SPA that somebody downloaded off
- 11:26Speaker 2: of Napster in nineteen ninety nine.
- 11:28Speaker 1: The entire video is definitely worth checking out, and will
- 11:31Speaker 1: include a link to it in the show notes. But
- 11:33Speaker 1: it's really interesting to hear how confused SUNO gets when
- 11:37Speaker 1: the track is encoded with Poisonify.
- 11:39Speaker 2: In some of those demonstrations, a lot of people were like,
- 11:41Speaker 2: that is so crazy, and it's like, no, really, what
- 11:43Speaker 2: it's doing is like, that's its own safety mechanism, Like
- 11:46Speaker 2: because it knows that it was confused.
- 11:49Speaker 1: This is pretty fascinating Poisonify. It doesn't make AI think
- 11:53Speaker 1: the drums or flutes, It just confuses it so much
- 11:57Speaker 1: with the noise that it adds to the spectrogram that
- 11:59Speaker 1: the AI doesn't know what to do and falls back
- 12:02Speaker 1: and randomly chooses something that it knows to be music.
- 12:06Speaker 2: So if you've very used like generative AI with images,
- 12:08Speaker 2: you'll notice something that happens quite often. As you'll say
- 12:10Speaker 2: I want a dog in a canoe eating a banana
- 12:14Speaker 2: headed towards a sunset, and the image you would get
- 12:19Speaker 2: might be like a dog on a jet ski not
- 12:23Speaker 2: eating anything in a lake with no sunset, And those
- 12:26Speaker 2: are really just failsafes like it tried to make a sunset,
- 12:29Speaker 2: it didn't have enough confidence. It literally is called the
- 12:31Speaker 2: confidence rating, and so it just said, okay, let's just
- 12:34Speaker 2: make what we normally make in the background. And so
- 12:37Speaker 2: it's sort of the same thing with music.
- 12:39Speaker 1: And there's another program that takes it a step beyond
- 12:41Speaker 1: poison Ifi. Instead of masking the instruments, it masks the
- 12:45Speaker 1: music itself.
- 12:47Speaker 2: Harmony Cloak. That's like above my pay grade. I don't
- 12:49Speaker 2: really understand how that model works. I'm just really glad
- 12:52Speaker 2: that they're working on it. But yeah, I mean they
- 12:54Speaker 2: obfuscate melody and harmony, which is pretty crazy.
- 13:00Speaker 1: I talked to the developer of Harmony Cloak about how
- 13:03Speaker 1: exactly they do this, and they even helped us test
- 13:06Speaker 1: it out on the kill Switch theme song that's after
- 13:09Speaker 1: the break. At the same time that Ben was working
- 13:24Speaker 1: on Poisonify, researchers at the University of Tennessee Knoxville were
- 13:28Speaker 1: working on another way to poison pill music, a program
- 13:32Speaker 1: they call Harmony Cloak.
- 13:34Speaker 4: Hu mess and machinks interpret data in different ways, so
- 13:39Speaker 4: there's a perceptual gap between humans and the machinks.
- 13:42Speaker 1: Jin Lu is an assistant professor at the University of Tennessee,
- 13:45Speaker 1: Knoxville and the lead developer of Harmony Cloak. He's also
- 13:49Speaker 1: really into music himself.
- 13:51Speaker 4: I love Milzaic. Actually I also play Militzach. I played
- 13:54Speaker 4: best guitar.
- 13:55Speaker 1: Harmony Cloak is similar to Poisonify and then it adds
- 13:58Speaker 1: imperceptible noise to the fire file. But unlike PUISONI five,
- 14:03Speaker 1: Homony Cloak doesn't just work on the level of the instruments.
- 14:06Speaker 1: It completely confuses the AI, so it can't learn from
- 14:09Speaker 1: the music at all.
- 14:11Speaker 4: So what we are doing right now is to use perturbation.
- 14:15Speaker 4: So we inject imperceptible perturbations to the music sampos to
- 14:21Speaker 4: trick the model into believing that they have already learned
- 14:25Speaker 4: this before. So there's no new knowledge, no new information
- 14:28Speaker 4: in bad in this music samples, so they couldn't learn
- 14:31Speaker 4: anything from this piece of work.
- 14:34Speaker 1: So the AI thinks there's no new information and essentially
- 14:38Speaker 1: ignores everything that's in the file. This means that AI
- 14:41Speaker 1: models can't train on music. With Harmony Cloak applied, you're
- 14:45Speaker 1: talking about introducing noise or what you call, you know,
- 14:50Speaker 1: technical term perturbations, right perturbations into the music, which you
- 14:54Speaker 1: say are imperceptible? Are these actually imperceptible? Right? If you're
- 14:59Speaker 1: adding noise, if you're adding extra data into the music,
- 15:03Speaker 1: can I as a listener hear.
- 15:05Speaker 4: That the perturbation we injected should have a minimum impact
- 15:10Speaker 4: on the perceptual quality of the music, because no one
- 15:14Speaker 4: wants to add noises to their artwork. So we conducted
- 15:18Speaker 4: a very comprehensive under study. We present both original one
- 15:22Speaker 4: and the perfect one to musicians and we ask them
- 15:25Speaker 4: to tell the difference, and our studies shows that they
- 15:29Speaker 4: won't tell difference between these two. I think in terms
- 15:32Speaker 4: of the musical quality, there's no big difference. So actually
- 15:36Speaker 4: the noises.
- 15:37Speaker 1: Itself is audible.
- 15:39Speaker 4: So if you listen to the noises only, like you
- 15:43Speaker 4: separate the perurbation from the music examples, you can hear
- 15:46Speaker 4: something it's audible. But if you combine these too, the
- 15:50Speaker 4: noises will be hidden under the music samples because we
- 15:54Speaker 4: leveraged the psycho acoustic phenomenon. So when we listen to
- 15:58Speaker 4: these music samples and pation and together, now is this
- 16:01Speaker 4: will become imperceptible.
- 16:03Speaker 1: I was curious about this whole imperceptible thing. So Jin
- 16:07Speaker 1: said he would not only help us test it, but
- 16:09Speaker 1: also use an updated process they're calling music shield. So
- 16:13Speaker 1: I'm gonna play a snippet of the kill Switch theme
- 16:15Speaker 1: song with and without music shield applied, and you see
- 16:19Speaker 1: if you can tell the difference. Here's sample one, okay,
- 16:57Speaker 1: and here is sample two, So you tell me. Could
- 17:37Speaker 1: you hear the difference? It's very slight, if anything. Oh
- 17:41Speaker 1: and by the way, if you were wondering, the one
- 17:43Speaker 1: that was run through music shield was the first sample.
- 17:47Speaker 1: And here's something else. Music shield actually goes one step
- 17:51Speaker 1: further than the harmony cloak process that we were talking about.
- 17:54Speaker 1: It not only stops AI models from training on the music,
- 17:58Speaker 1: but it can also prevent music generators like Suno or
- 18:01Speaker 1: Udio from being able to edit or remix tracks. So
- 18:05Speaker 1: let's give it a shot, and let's put this thing
- 18:07Speaker 1: into Suno. So if we upload our original untreated theme
- 18:11Speaker 1: into Suno and tell it to remix it, here's what
- 18:14Speaker 1: we get, which for an AI generator is not bad.
- 18:52Speaker 1: That's in the ballpark of what the original song sounds like.
- 18:55Speaker 1: And this is what happens when we upload our same
- 18:58Speaker 1: theme that's been music shielded. Okay, Yeah, this is different.
- 19:21Speaker 1: It's a lot more soothing than the song we gave it.
- 19:26Speaker 1: It kind of feels like a corporate video, maybe for
- 19:29Speaker 1: investors at a defense contractor. There's maybe a guy in
- 19:33Speaker 1: a suit on the screen and he's telling you about
- 19:35Speaker 1: how their business is really all about family. Clearly it works,
- 19:43Speaker 1: Zuno got so confused that it just spit out some
- 19:46Speaker 1: generic corporate music. So I've read your paper that you
- 19:50Speaker 1: published recently entitled Harmony Cloak Making Music Unlearnable for Generative AI.
- 19:57Speaker 1: One of the things that I found really interesting is
- 20:01Speaker 1: the language that you use, right, because you're osensibly you're
- 20:05Speaker 1: talking about music, very kind of broad, very easy to understand. Thing.
- 20:11Speaker 1: Because I'm reading through your paper, I'm realizing I'm reading
- 20:15Speaker 1: a security paper. Section three point one is entitled threat model.
- 20:21Speaker 1: And then as I read further down, I'm seeing, you know,
- 20:23Speaker 1: I'm just gonna read from this a little bit. Yeah,
- 20:25Speaker 1: you're laughing, but this is amazing to me. I mean,
- 20:28Speaker 1: the attacker eg AI companies or model owners might scrape
- 20:32Speaker 1: music data from the Internet or music streaming platforms to
- 20:35Speaker 1: train their music generative AI models, potentially leading to copyright
- 20:39Speaker 1: infringements and harming musicians. This part right here, I love
- 20:43Speaker 1: this we assume the attacker possesses substantial advantages and capabilities,
- 20:48Speaker 1: including unrestricted access to the training data set and model parameters,
- 20:52Speaker 1: facilitating comprehensive data and grading expansions, and the ability to
- 20:56Speaker 1: perform adaptive attack strategies. And it goes on. But I mean,
- 21:01Speaker 1: this is fascinating because usually if you read it security paper,
- 21:04Speaker 1: you're thinking of the defender is you know, somebody with
- 21:08Speaker 1: some resources. It could be a bank, it could be
- 21:09Speaker 1: a tech company, it could be a governmental agency, and
- 21:13Speaker 1: the attacker is somebody with considerably less resources. This is
- 21:16Speaker 1: the reverse of that. The attacker is somebody with a
- 21:20Speaker 1: lot of resources, it's a large tech company probably, and
- 21:24Speaker 1: the defender is just you know, some kid with the
- 21:26Speaker 1: bass guitar.
- 21:28Speaker 4: Yeah, exactly. And actually in the paper we also discussed
- 21:31Speaker 4: the possible a tech because the big tech company may
- 21:35Speaker 4: also leverage additional strategy to relearn or process. You are
- 21:40Speaker 4: protecting music to learn something from it. Right, So one
- 21:45Speaker 4: very straightforward way is to use noise cancelation techniques to
- 21:50Speaker 4: remove any perturbations from the music samposts. So that's maybe
- 21:56Speaker 4: one strategy they leverage. But the easy here is that
- 22:01Speaker 4: you can leverage whatever way to remove noises. Yes, you
- 22:05Speaker 4: can reduce the effectiveness of our framework, but on the
- 22:10Speaker 4: other side, the quality of the music will be dropped
- 22:13Speaker 4: as well, because when you remove noise is certain music
- 22:16Speaker 4: or futures will be removed as well.
- 22:19Speaker 1: Yeah, so I think I see what you're saying here.
- 22:21Speaker 1: So your framework, harmony cloak, the entire purpose is to
- 22:26Speaker 1: make the song unusable for the tech company. And if
- 22:29Speaker 1: the tech company then tries to do something to mitigate that,
- 22:33Speaker 1: to remove the noise, the perturbations that you've introduced into
- 22:37Speaker 1: the music, that reduces the quality, and do they really
- 22:40Speaker 1: want to be putting bad quality data into the data set. No,
- 22:44Speaker 1: So you've accomplished your goal, which is again to make
- 22:47Speaker 1: it unusable for them.
- 22:48Speaker 4: Yeah.
- 22:49Speaker 1: Ben Jordan has a similar philosophy with this Poisonified project.
- 22:53Speaker 1: He knows it's going to be a battle with AI companies,
- 22:56Speaker 1: but that's kind of the point.
- 22:58Speaker 2: A lot of people have said, well, you know, with
- 23:00Speaker 2: what happened with Nightshade and Images, they're just going to
- 23:03Speaker 2: do something like that.
- 23:05Speaker 1: To clarify what Ben's talking about here. Pretty soon after
- 23:08Speaker 1: Nightshade came out, and that's the poison pilling tool for
- 23:11Speaker 1: Images mentioned earlier people started saying that they'd figured out
- 23:14Speaker 1: a way to bypass it by blurting or sharpening out
- 23:17Speaker 1: the noise. The team that developed Nightshade disputes this, but
- 23:21Speaker 1: their other popular tool called Glade, was also briefly bypassed
- 23:26Speaker 1: using some image upscaling techniques. And it's basically a back
- 23:30Speaker 1: and forth war here. And because with audio AI still
- 23:34Speaker 1: processing a spectrogram image, in theory an AI company could
- 23:39Speaker 1: also use similar techniques to bypass poisonify and harmony cloak.
- 23:43Speaker 2: They might There are a couple considerations out, So like
- 23:46Speaker 2: when we talked about that audio going into a spectrogram
- 23:48Speaker 2: back into audio thing, if you think about what would
- 23:51Speaker 2: happen if you think about how a snare drum works
- 23:54Speaker 2: and a spectrogram, Okay, it's like it kind of starts
- 23:56Speaker 2: immediately and then it fades out a little bit, and
- 23:59Speaker 2: that gives it like the right, Yeah, if you were
- 24:03Speaker 2: to bore that, now it sounds really bad.
- 24:07Speaker 1: You have this.
- 24:10Speaker 2: And so different things need to be precise. And not
- 24:14Speaker 2: only that, so they use like the boring and the
- 24:15Speaker 2: AI sharpening, and if they were to do that with spectrograms,
- 24:18Speaker 2: it still that's a lot of extra compute and expense,
- 24:23Speaker 2: and really the goal is to just pressure them to
- 24:26Speaker 2: work with musicians, Like if they actually want to make
- 24:28Speaker 2: money off of this and they want to continue selling
- 24:31Speaker 2: subscriptions to generate this stuff, then uh, you know, it's
- 24:35Speaker 2: just to pressure them to make getting the wavefile directly
- 24:39Speaker 2: from a musician cheaper and easier than doing it the
- 24:42Speaker 2: way they've been doing it without consent.
- 24:44Speaker 1: Uh huh. So part of this is you're not necessarily
- 24:48Speaker 1: thinking that this is an undefeatable attack against you know,
- 24:54Speaker 1: AI scraping. It kind of sounds like you're sort of
- 24:57Speaker 1: hoping that it becomes obsolete because we know there's going
- 25:00Speaker 1: to be an arms race. We know companies are going
- 25:03Speaker 1: to figure out a way to defeat your poison pill attack.
- 25:08Speaker 1: Let's just be real. They got more resources than you,
- 25:10Speaker 1: they got more engineers than you do. Yeah, they're going
- 25:12Speaker 1: to figure it out, sure, but it kind of sounds
- 25:14Speaker 1: like you just rather them decide, you know what, the
- 25:16Speaker 1: sain't worth it?
- 25:17Speaker 2: Yeah, I mean you know what I do. Hear like
- 25:20Speaker 2: the arms race analogy all the time, and it's like
- 25:22Speaker 2: war is almost always a net loss, Like and if
- 25:26Speaker 2: they have anybody smart in whoever's funding them, they won't understand.
- 25:30Speaker 1: That you can make it so annoying, yeah, to scrape
- 25:34Speaker 1: people's music that they're just going to go the you know,
- 25:38Speaker 1: quote unquote do the right thing.
- 25:39Speaker 2: Yeah, Ultimately you might have it to be is so
- 25:42Speaker 2: omnipresent that AI music sites actually have to say, okay,
- 25:46Speaker 2: well we need to just talk to artists now, we
- 25:48Speaker 2: need to just start training on stuff that we know
- 25:50Speaker 2: doesn't have this, because we're wasting too much money on
- 25:53Speaker 2: compute for things that are just degrading the model quality.
- 25:59Speaker 1: So if you're a musician, you're probably thinking, when can
- 26:03Speaker 1: I start using this stuff to protect my music? Or
- 26:05Speaker 1: if you're a music fan, you might want to know
- 26:07Speaker 1: when this stuff drops so your favorite artists can stop
- 26:10Speaker 1: getting their music scraped and stolen. Well, I've got bad
- 26:14Speaker 1: news for you, but also some good news that's after
- 26:17Speaker 1: the break. So I know that there's gonna be definitely
- 26:34Speaker 1: some musicians who will hear this and say this sounds amazing.
- 26:38Speaker 1: I want this now? Yeah, is this something that's available
- 26:41Speaker 1: right now? Now?
- 26:43Speaker 2: It's funny because after I release that video, I probably
- 26:46Speaker 2: got one hundred emails of people just linking me to
- 26:48Speaker 2: like Google drive of their songs and I'm just like, okay,
- 26:51Speaker 2: this is not how it works, like unfortunately, yeah, so
- 26:55Speaker 2: I am not like, I'm not an mL developer. I
- 26:58Speaker 2: can't write a phision code. I believe it took about
- 27:02Speaker 2: two weeks for ten songs or something like that on
- 27:06Speaker 2: my machine with two brand news state of the art,
- 27:10Speaker 2: you know, big video cards.
- 27:11Speaker 1: Two weeks as in on and off incoding, no NonStop
- 27:15Speaker 1: two weeks and non stop and coding to do your album. Okay, yeah,
- 27:17Speaker 1: that is that's not accessible. I will not be asking
- 27:20Speaker 1: you to handle my album for me then never mind.
- 27:22Speaker 2: Yeah, And so it's also like even if I could,
- 27:24Speaker 2: then it's still like, okay, well if it's this inefficient,
- 27:26Speaker 2: like using this much power, you know, what we don't
- 27:30Speaker 2: want is to set the planet on fire just to
- 27:33Speaker 2: protect our music from a couple startups.
- 27:37Speaker 1: So Ben Jordan's program might not be available to the
- 27:40Speaker 1: public anytime soon. That's the bad news. But here's some
- 27:43Speaker 1: good news. Professor Jin Leeu and his team do have
- 27:46Speaker 1: some near future plans to make their software more widely
- 27:49Speaker 1: available for a musician in the future. What would protecting
- 27:55Speaker 1: your music with something like Harmony Cloak look like? Is
- 27:59Speaker 1: it downloading an app? Is it, uploading it to a
- 28:02Speaker 1: site and redownloading it. What are they doing?
- 28:05Speaker 4: There are many many ways to use this technology to
- 28:08Speaker 4: protect their music. First of all, we are thinking to
- 28:12Speaker 4: integrate these technologies with other platforms, for example, Apple Music Spodify,
- 28:19Speaker 4: so in that case, once they upload their music to
- 28:22Speaker 4: their platform, they can automatically protect their music. We'll also
- 28:27Speaker 4: create a web set, so on our web set they
- 28:30Speaker 4: can upload their music then download the perturbrization version from it,
- 28:35Speaker 4: so musicians people can use this very easily.
- 28:37Speaker 1: Do you have a timeline for when this might be
- 28:39Speaker 1: available for the public.
- 28:41Speaker 4: In July, we plan to launch a test program which
- 28:44Speaker 4: will involve around two hundred musicians so that we can
- 28:47Speaker 4: folder fun team fold improved this system before large scale deployment,
- 28:53Speaker 4: and if everything goes smoothly, I think integration of this
- 28:59Speaker 4: technology in at the Plasma will be very quick. Hopefully
- 29:02Speaker 4: this can be integrated in August ord September this year.
- 29:09Speaker 1: Despite the fact that he's working on programs that are
- 29:11Speaker 1: actively fighting against AI, jin, Lu is not universally anti AI,
- 29:17Speaker 1: and neither is Ben Jordan. They both think that AI
- 29:20Speaker 1: can be a useful tool.
- 29:22Speaker 4: I think AI machine learning itself doesn't have any problems.
- 29:26Speaker 4: The problem is how this big tech company trend their models.
- 29:30Speaker 4: And also from the musician's perspective, because we talk to
- 29:34Speaker 4: many many musicians, actually some of them use AI. They
- 29:37Speaker 4: feel this AI model is pretty useful. But if these
- 29:41Speaker 4: company wants to use their music examples for training models,
- 29:44Speaker 4: they need to get explicit permission. Also they need to
- 29:49Speaker 4: offer compensation to musicians.
- 29:54Speaker 1: How do you feel like, say the next six months
- 29:56Speaker 1: year plays out for music and AI.
- 30:00Speaker 2: One thing that I mean it's probably good news for
- 30:04Speaker 2: anybody who's worried about AI music taking over anything, is that,
- 30:07Speaker 2: like psychoacoustics are really really complicated, Like telling a computer
- 30:13Speaker 2: to hear something without any sort of image analysis, to
- 30:16Speaker 2: just not go that spectral conversion route and just hear
- 30:19Speaker 2: something the way a human hears is like you may
- 30:22Speaker 2: as well just ask it to become self aware, because
- 30:25Speaker 2: that sounds easier to me, just because like what's happening,
- 30:29Speaker 2: you know, like our hearing is by far our most
- 30:31Speaker 2: sensitive sense that we have. And when you think about
- 30:35Speaker 2: what happens from picking up pressure waves and then little
- 30:39Speaker 2: hairs in our ear picking this up and interpreting them
- 30:42Speaker 2: in conjunction with our brain into sounds. It's kind of
- 30:46Speaker 2: mysterious and crazy. And so to just tell AI, like, hey,
- 30:51Speaker 2: listen to this this sonic pressure and figure out how
- 30:55Speaker 2: to make it again like that, that's a much bigger
- 30:57Speaker 2: ask than I think it sounds too your average investor
- 31:01Speaker 2: or something who thinks that AI music is going to
- 31:04Speaker 2: eventually I don't know, I guess replace musicians or something.
- 31:09Speaker 1: Yeah.
- 31:09Speaker 2: Ideally, what I would really like to see is, as
- 31:13Speaker 2: people get more used to AI, two things I would
- 31:15Speaker 2: like people to use it locally. So, for example, image
- 31:19Speaker 2: and heap. She sent me a bunch of like probably
- 31:21Speaker 2: over an hour of her singing, sometimes in like really
- 31:24Speaker 2: weird ways, and then we sort of work together and
- 31:26Speaker 2: I create a voice model, and then she sang through
- 31:28Speaker 2: the voice model, and that was all locally. It wasn't
- 31:30Speaker 2: happening through any sort of service. I really liked that idea.
- 31:33Speaker 2: And I like the idea of artists being able to
- 31:37Speaker 2: sell their voice or sell their music style or their
- 31:40Speaker 2: instruments or something like that and put it in a marketplace.
- 31:43Speaker 2: And really the only technology that would need to exist
- 31:47Speaker 2: in any sort of centralized way would just be somebody
- 31:49Speaker 2: to water market or something. I really like that and
- 31:52Speaker 2: the other thing is, right now, we're in this land
- 31:55Speaker 2: in generative AI where the ideas are just huge and
- 32:01Speaker 2: kind of nonsensical, Like, you know, what if AI replaced music?
- 32:05Speaker 2: It's like, Nope, it's not gonna do that, But what
- 32:08Speaker 2: if it replaced samplers, you know, like violin, sample instruments
- 32:13Speaker 2: or something like. That's actually somewhere where AI can do
- 32:15Speaker 2: a really really good job to make writing music more
- 32:19Speaker 2: fun and more accurate, I guess, and you know, things
- 32:22Speaker 2: like that. And so once we sort of realize that
- 32:25Speaker 2: that not every single person is going to adopt AI
- 32:28Speaker 2: music and stop listening to humans, then maybe we can
- 32:32Speaker 2: invest money into making practical solutions.
- 32:38Speaker 1: And that is it for this particular discussion about AI
- 32:42Speaker 1: and music. And I say this particular discussion because this
- 32:45Speaker 1: is not the last time we're going to be talking
- 32:47Speaker 1: about AI and music. This is a really big topic
- 32:50Speaker 1: and all of us at kill Switch are pretty into music,
- 32:53Speaker 1: so we're absolutely going to be getting back into this again.
- 32:56Speaker 1: And you know, if you've got any music related stuff
- 32:58Speaker 1: that you're curious about, let us No. Before I get
- 33:01Speaker 1: out of here, though, I gotta do some shout outs. First,
- 33:04Speaker 1: big shout out to Ben Jordan. If you found his
- 33:06Speaker 1: Poisonify concept interesting, he has a whole YouTube video on it,
- 33:10Speaker 1: and the link for that is in the show notes.
- 33:12Speaker 1: He's also started a company called top Set Labs that's
- 33:16Speaker 1: developing AI voice models that are trained on artists who
- 33:19Speaker 1: have given their explicit consent, and a lot of them
- 33:22Speaker 1: are making more money in those royalties than they do
- 33:24Speaker 1: on Spotify. So if you're a musician or you're just
- 33:27Speaker 1: curious on how that works, might want to check that
- 33:30Speaker 1: out too. Also a big shout out to our other guest,
- 33:32Speaker 1: professor jenlu as well as Saydeerfond from the University of
- 33:36Speaker 1: Tennessee Knoxville for letting us test out Harmony Cloak and
- 33:39Speaker 1: music Shield. And if you want to check out that
- 33:41Speaker 1: paper we were referencing, there's a link to that also
- 33:43Speaker 1: in the show notes. Thank you so much again for
- 33:46Speaker 1: listening to kill Switch and again let us know what
- 33:49Speaker 1: you think and if there's something you want us to cover,
- 33:51Speaker 1: We're easy to find. You can hit us up at
- 33:53Speaker 1: kill Switch at Kaleidoscope dot NYC, or you can check
- 33:57Speaker 1: us out on Instagram at kill switch pod or I'm
- 34:00Speaker 1: dex Digi. That's d e x d I g I
- 34:03Speaker 1: on Instagram or Blue Sky, and wherever you're listening to us,
- 34:07Speaker 1: make sure to leave us a review because it helps
- 34:09Speaker 1: other people find the show and that helps us keep
- 34:12Speaker 1: doing our things. Kill Switch is hosted by Me Dexter Thomas.
- 34:15Speaker 1: It's produced by Shina Ozaki, Darluk Potts and Kate Osborne.
- 34:20Speaker 1: Our theme song is by me and Kyle Murdoch and
- 34:22Speaker 1: Kyle also mixed the show. From Kaleidoscope. Our executive producers
- 34:26Speaker 1: are Ozo Lashin, Mangesh Hatikadur and Kate Osborne. From iHeart,
- 34:31Speaker 1: our executive producers are Katrina Norville and Nikki E. Tour
- 34:35Speaker 1: catch All the Next One