Shielded: The Last Line of Cyber Defense
Hardware Is the Hidden Risk: Ferhat Yaman on PQC, Side-Channel Attacks, and AI Privacy
August 7, 2025
What if your hardware is leaking secrets before your software even boots? In this episode of Shielded: The Last Line of Cyber Defense, Johannes Lintzen sits down with Ferhat Yaman, a security researcher at AMD’s Product Security Office, to explore the front lines of post-quantum cryptography, hardware vulnerabilities, and AI privacy. From electromagnetic side-channel attacks to homomorphic encryption, Ferhat shares practical insights from years of research and testing across AMD, academia, and open source projects. Learn what it takes to build and test hardware for a post-quantum world.
As post-quantum cryptography moves from theory to hardware, organizations can no longer afford to ignore the physical layer of security. In this episode of Shielded: The Last Line of Cyber Defense, host Johannes Lintzen speaks with Ferhat Yaman, security researcher at AMD’s Product Security Office, to explore how electromagnetic side-channel attacks, hybrid cryptography, and AI privacy are reshaping the future of secure system design. They discuss the risks of leakage in post-quantum implementations, the challenge of model theft in AI accelerators, and why mitigation needs to start before silicon is even taped out. From masking and shuffling to pre-silicon testing and homomorphic encryption, Ferhat offers a candid and deeply technical look at what it means to build quantum-resilient systems in hardware, not just in code.

What You’ll Learn:

Ferhat Yaman is a security researcher at AMD's product security office, where his work spans post-quantum cryptography, AI privacy, and side-channel resilience. With a background in both theoretical cryptography and practical hardware design, Ferhat has contributed to projects including the Crystals-Kyber and Dilithium PQC implementations, Caliptra Root of Trust, and electromagnetic model extraction from Google’s Edge TPU. His research explores how secure systems can be built from the silicon up, balancing performance, cost, and long-term quantum readiness. Ferhat’s recent work looks at accelerating homomorphic encryption for AI workloads and improving pre-silicon testing using commercial and open-source tools.

With the shift to post-quantum hardware security accelerating, Yaman’s message is clear: protecting systems requires more than new math; it demands early testing, layered defenses, and security built into the silicon itself.

Your Roadmap to Hardware-Centric PQC:

[06:59] Step 1: Test for Physical Leakage, Not Just Algorithm Strength - 

Strong algorithms don’t guarantee strong protection if the hardware leaks secrets. Ferhat explains how side-channel attacks can extract private keys by analyzing power consumption or electromagnetic emissions, especially in hardware implementations of post-quantum cryptography like Kyber and Dilithium. Even minor hardware optimizations meant to improve speed can unintentionally introduce new leakages. Key Question: Have you tested your PQC hardware for side-channel leakage, or just verified the math?

[11:19] Step 2: Prepare for Hybrid Attacks, Not Just Hybrid Crypto - 

Attackers are blending techniques, merging cryptanalytic insights with side-channel data to break even well-implemented systems. Ferhat emphasizes that defending against these multi-layered threats requires layered countermeasures across hardware and software. The combination of multiple attack vectors makes traditional assumptions about isolated vulnerabilities dangerously outdated. Key Question: Are you planning for real-world attack combinations or idealized test conditions?

[13:52] Step 3: Use Hybrid Cryptography to Bridge the Transition -

You don’t have to replace everything overnight. Ferhat describes how AMD’s Caliptra project integrates classical and post-quantum cryptography side-by-side, giving developers flexibility while building resilience. Hybrid cryptography is a practical way to support both legacy and quantum-safe systems, especially in constrained environments. It’s a crucial stepping stone for organizations that can’t afford a full cryptographic overhaul in one cycle. Key Question: Can your architecture support both today’s standards and tomorrow’s requirements?

[17:38] Step 4: Defend AI Models from Electromagnetic Model Theft - 

AI workloads running on edge hardware can leak critical information. Ferhat’s research shows how electromagnetic side-channel attacks can reveal the structure of neural networks down to the number of layers and nodes, without touching the model file. For companies that treat models as IP, this presents a new category of risk. This form of model extraction reduces training time for attackers by shortcutting hyperparameter tuning. Key Question: Could your AI accelerators be quietly leaking trade secrets?

[23:52] Step 5: Build Security In, With Countermeasures That Match the Threat -

Mitigation strategies like masking, shuffling, and randomness injection aren’t optional; they’re essential. Ferhat walks through how these hardware-level techniques help minimize leakage, but warns that each comes with design and performance tradeoffs. Choosing the right countermeasure depends on your timing, area, and risk budget; there’s no one-size-fits-all solution. Key Question: Have you budgeted for countermeasures, or are you counting on hope?

Episode Resources:

Want exclusive insights on quantum migration?  Stay ahead of the curve. Subscribe to Shielded: The Last Line of Cyber Defense on Apple Podcasts, Spotify, or YouTube Podcasts.

✔ Get insider knowledge from leading cybersecurity experts.


✔ Learn practical steps to future-proof your organization.


✔ Stay updated on regulatory changes and industry trends.

Need help subscribing? Click here for step-by-step instructions.

Shielded: The Last Line of Cyber Defense is handcrafted by our friends over at: fame.so

Transcript

Johannes 0:02 Today on Shielded, the last line of cyber defense. We are joined by Ferhat Yaman, a security researcher at AMD's Product Security Office whose cutting edge work sits at the intersection of hardware security, post quantum cryptography and AI privacy. As quantum computing rapidly approaches practical viability, organizations face a critical window to future proof their security infrastructure. Not only at the software level, but also deep inside the hardware that underpins AI and cryptographic systems. Fairhut's research including groundbreaking demonstrations of AI model extraction via electromagnetic side channel attacks on hardware accelerators. Now that's a mouthful. Reveals stark vulnerabilities that many organizations don't yet fully grasp. His expertise in securing hardware implementations of quantum resistant cryptography and protecting AI privacy offers crucial and actionable insights for those aiming to stay ahead of quantum threats. Quantum readiness isn't just about adopting new algorithms. It's about understanding the practical realities and risks involved in hardware deployments today. Ferhan, welcome to Shielded. It's a real pleasure to have you here.

Ferhat 1:22 Thank you for inviting me here, Joe. Nice to meet you.

Johannes 1:26 Absolutely. My pleasure. So, so Ferhat, I mean, originally you are from Istanbul, Turkey. But I understand you call Boston home now. Now I want to know what time of year is it that you miss Turkey the most?

Ferhat 1:41 Yeah, I guess. Yeah. It's the spring is the best time, uh, in the Turkey. So yeah, it's, don't have my dad call. Everything is flowing. Even then, Istanbul, like, you have the Turkish daffy and teas and the little Indianites and then the Beijing side. It's probably messing up.

Johannes 2:02 That's lovely. Yeah. So that's the fine ground stuff, right, that you you cook.

Ferhat 2:06 And then

Johannes 2:06 it accumulates at the bottom. It's like, right? Yeah. Nice and sweet. Right. So besides coffee and the and the climate, I'm I'm sure that, uh, the Turkish cuisine also plays a role because uh, some fantastic food options in, in Istanbul and, and around.

Ferhat 2:23 Yeah. So that is the, yeah, I mean, um, relation with the mother, so always, you know, still like Turkish food. And my wife and I always try to like bring that home side of it, like, um, deliciousness. But yeah, we are always missing the, the, the, like the essence of it.

Johannes 2:47 Yeah. I can imagine. I can imagine. So besides your, you know, geographical journey, of course, your career journey is fascinating as well. You know, going from academia to now AMD, always with a focus on side channel and analysis and and post quantum cryptography. Could you start by by sharing how your background and research on hardware vulnerabilities and AI privacy really intersect with the urgent challenges that organizations are facing today as they hopefully prepare for post quantum cryptography? with a focus on side channel and analysis and and post quantum cryptography. Could you start by by sharing how your background and research on hardware vulnerabilities and AI privacy really intersect with the urgent challenges that organizations are facing today as they hopefully prepare for post quantum cryptography?

Ferhat 3:21 Yeah. I actually started working on this, like, almost more than eight years ago. So, yeah, it was, like, more on the, like the AI privacy side of it, but it actually introduced me a lot of on the, like the Lattice crypto. So this is like one of the main, uh, building blocks in the, um, post quantum cryptography. But it was also used in the, like, AI privacy, which is called, like, homomorphic encryption. So that basically also allows us to, like, uh, do some, uh, processing and, uh, on the, like, the data without decrypting it. So that is also something that that is, uh, enormous for us, actually. So that was I I was started, like, eight years ago, and then it just, like, pulled me in this. So yeah. Then I see that, like, okay. We are doing some processing in in, like, we're not decrypting anything, and we we are not seeing any inf like, confidential data. And then it moved into the Lattice crypto. Once I started working on this, and then I realized the existence standardization is going on with the, like, NIST, and then I just continue working on this, basically. Um, Yeah.

Johannes 4:40 So really those those things culminating, coming together?

Ferhat 4:44 Yes. And then Yeah. Yeah. Yeah. It it was, like, after, like, on this, um, after when I realized there is this, um, NEST standardization. So that was a postdoc in back in Turkey, my university while I was undergrad. So, um, he was kind of invited, okay, you are working on this latest crypto side of it. So, okay, why we are not focusing on this, uh, post quantum crypto and there is this crisis bridge, so which is quite similar. And then since I have some kind of hardware design part of it, and we start moving on like, okay, how we can optimize things and how we can, like, kind of leverage the the hardware design or, like, to introduce some kind of small optimizations on on the ArtMagic stuff of the lattice script. Right. So and, uh, we we were actually was able to design a new KFAST on manual multiplication hardware for the crystal squares. So it was a kind of, um, like, kind of at the time because that was a kind of there's a, like, module reduction units and then the PQC that you are basically you have to, like, design specifically for crystal sky bridge. So that basically leverage our design and then then more into it. But, yeah, that was the kind of the, um, more, uh, nuance on on this work, actually. because that was a kind of there's a, like, module reduction units and then the PQC that you are basically you have to, like, design specifically for crystal sky bridge. So that basically leverage our design and then then more into it. But, yeah, that was the kind of the, um, more, uh, nuance on on this work, actually.

Johannes 6:16 The new new new new ones in in, uh, in in the lattice based world. Yeah. And and, of course, you know, I mean, we're talking about side channel attacks, and, of course, they're nothing new. Right? I mean, they've they've always been around. They give long post risks to to any type of cryptographic implementation in in the classical world and and just the same in in, you know, the new lattice based, uh, um, post quantum world. So I think, you know, as PQC algorithms begin to more broadly be deployed in hardware, what are some of the unique challenges and vulnerabilities that are introduced by those, um, quantum resistant, um, cryptographic designs in in hardware.

Ferhat 7:00 Yeah. I mean, um, one of like, there are a lot of things, like, if you implement those algorithms in like, software, there are different challenges. If you're designing hardware, it is always a different challenge. So you have to design in a way that you're not leaking anything regarding a secret key or any information regarding the operations on as well. So my focus was on more on the, like, side channel analysis of those. It was classic crypto as well, and AEs. So I was also most mostly doing those rigs before, like, studying and dealing in here as well. So is mostly looking for. If there is a, like, a key processing happening, like, secret key processing or execution happening during the operations, is there some some kind of, like, side channel we can language it? So, mostly, we are looking into the physical side channels, like our construction of the device or electromagnetic radiations coming from the device. And we are looking of those, like, traces, like we call this, of and electromagnetic radiations, and then try to kind of correlate, um, interest some kind of hypothetical model we have, and then try to to find a way to speak out of it. That that is the objective one. In the features yeah. So you

Johannes 8:41 look so you look for a specific pattern that looks like that could be, uh, you know, the same format as you'd expect

Ferhat 8:48 that he to correct. Yeah. Yeah. It's mostly for for example, um, uh, for let's say, like in the, let's say, in the the algorithm, which is the, like, signature at

Johannes 9:05 DSA. MLDSA. Yeah. r a specific pattern that looks like that could be, uh, you know, the same format as you'd expect

Ferhat 8:48 that he to correct. Yeah. Yeah. It's mostly for for example, um, uh, for let's say, like in the, let's say, in the the algorithm, which is the, like, signature at

Johannes 9:05 DSA. MLDSA. Yeah.

Ferhat 9:07 Yeah. This is the one that generally, like, focused on. So for example, in the key key generation and then as well as the, like, signature part of it. So there is this key operations happening. So you can basically, um, pivot your, like, execution and doing the whole number multiplication. And, actually, that secret key is generated or secret key is getting executed to generate the signature. So you can basically see when it starts and when it ends. And based on the the multiplication operations that can use in in in between, so you can kind of design the power model and based on this. So it is also important that, like, if we are actually, like, introducing some kind of small optimization, this will also, like, kind of lead, um, some, uh, different kind of leakages or faulty parts as well.

Johannes 10:11 So optimizing for for speed, for instance, could could lead to, um, you know, spikes in power consumption. And then,

Ferhat 10:20 uh, yeah. That is correct. So for example, a recent work in this one is find out that, like, if we design a specific module reduction of good things. So for the Latium, we may figure out there is some part of the the addition unit of the mod like the module reduction part. So you can kind of pivot that specific part, and you can basically measure the, like, what kind of key is used during that operation, basically. Yeah. That was the recent work on this one.

Johannes 11:01 Got the very yeah. So so, I guess, um, as much as the protection mechanisms need to change, do you also see that the the side channel attacks itself? Do they need to adapt and and evolve in in this given context as well?

Ferhat 11:19 Yeah. I mean, you can combine different attacks, definitely. You can, uh, also, like, leverage the post quantum algorithm solutions as well. So, like, for example, one of the attacks is, like, also using, um, like, crypto analytic part of post quantum cryptography. And, also, some of them is using the side channel analysis and the obtained information from them. And then they combine both of it, and then they can kind of design a new way of attack of this one. So this was the, a couple of years ago, that was the the work that that been done actually.

Johannes 12:05 And and in that in that research, um, did you find anything unexpected, like, uh, you know, looking for specific vulnerabilities in in those implementations of of quantum resistant hardware design? his one. So this was the, a couple of years ago, that was the the work that that been done actually. And and in that in that research, um, did you find anything unexpected, like, uh, you know, looking for specific vulnerabilities in in those implementations of of quantum resistant hardware design?

Ferhat 12:19 Uh, I personally, um, I recently started working on the, like, open source algorithm, the culture, which is a autofrust algorithm, like, design and, like, design developed by AMD, Microsoft, Nvidia, etcetera, consortium to to develop this, uh, root of trust, uh, hardware design. It's basically involved a lot of classic crypto as well as the post quantum crypto in both of them. I was recently started looking at it. I did not show, like, final leakage so far, but there are some mechanisms that shows that there was a reason that work on this one done by research is from, I think, from the I think what the university is, so that they were, um, figure out the way of how can they disclose the key. We will also look into this at hand. Time. I think we we can try to duplicate and then if we can improve on it. So we'll we'll look at so we'll look at into that. Okay. So

Johannes 13:39 So so you said, um, with Calyptra, um, it's using classical as well as as post quantum algorithms. So so is it is it using hybrid cryptography then?

Ferhat 13:53 Yeah. It is actually, um, like, providing providing the bot options to the, like, the users. Yeah. It's like like like, as you can see that the the post quantum cryptography deployment is also kind of a different thing. You have to, like, kind of provide one of the users that you have this classic crypto still in place. And then, okay, we are start starting, uh, implementing more secure ones. So we have this option as well. So, like, you can have to give that option to the users as well. So that is the idea on this.

Johannes 14:30 So then I I can imagine on sort of the hardware level with, like, limited resources being available and, you know, size and, uh, size requirements to come in a topic as well?

Ferhat 14:45 Yes. Yes. I mean, um, that will be, like, important, like, how you can combine and then how you can also, like, provide both of them in the same platform. So those are going to be very important. And then, yeah, it is a kind of a a regulation game. Yeah. We can understand that when I think that was one of your previous podcast you're discussing about. So, um, uh, but, yeah, definitely for hardware as well, we have to have this option and availability for this

Johannes 15:23 Right.

Ferhat 15:23 Postcode encrypt algorithms.

Johannes 15:25 No. That makes sense. Yeah. So so I guess, you know, in in the experience that you've you've gained over the years working on those hardware security research projects at AMD, um, what are some of the main critical practical steps that, uh, you know, if I'm an organization, like, where should I prioritize in, in my PQC hardware implementation? I'm making sure that I'm resilient, um, ideally to again, both classical and, and, you know, quantum era science. e security research projects at AMD, um, what are some of the main critical practical steps that, uh, you know, if I'm an organization, like, where should I prioritize in, in my PQC hardware implementation? I'm making sure that I'm resilient, um, ideally to again, both classical and, and, you know, quantum era science.

Ferhat 16:00 Yeah. In AMD perspective, we generally deploy those algorithms in, um, in, um, mostly in book building. The bootloader will feel like whenever that there's a secure boot of it. So we should have to make sure that everything is verified, uh, correctly.

Johannes 16:20 You bring up the whole stack.

Ferhat 16:22 Correct. Yeah. Operating system is or or, like, other firmware components we have to verify and validate before it start loading it and executing it. So, currently, yeah, we are using QuestCrypto, and and then we are actually plan to play moving forward with the deployment of post quantum crypto with the and the others as well. So so this is our intention. I prescripted on this. Um, but, yeah, we'll we'll actually use this, like, this use case. And also, there are different use cases. You can also employ this post content cryptography, um, in in mostly even when you can just kind of put your own secure processor so that your operating system also can start using it. You are, like, in safety or, like, start to I tend to could also use it in this kind of using installation. So for for some of the notes you're doing it. So that is also going to be crucial. Yes. We are looking into it.

Johannes 17:39 Um, and and part of the work that you've done, um, was also together with the team, um, exposing the TPU extract attack on on, uh, I think it was Google's Edge TPU. Um, and then, of course, was a revelation on how electromagnetic I think you used the term EMF earlier. Right? So electromagnetic, uh, side channel analysis can can leak AI model parameters. Maybe maybe you could explain why this type of attack also matters for we were talking about AI privacy. Right? And what this implies about the security of AI accelerators.

Ferhat 18:17 Yes. Yeah. This was the branch I started doing my masters before joining the AMD. So this is regarding the work. So the important part of it okay. We can do this side channel attacks and cryptographic algorithms so that we can give you all the secret queue on this one. What okay. What if we use this attacks on on machine learning algorithms? Or, like, can we review, um, like, some some of the machine learning models, which is important for some of the companies because they're spending lots of hours to train those, and then it becomes kind of a IP or a way to use it.

Johannes 19:09 The secret sauce. Yeah. s attacks on on machine learning algorithms? Or, like, can we review, um, like, some some of the machine learning models, which is important for some of the companies because they're spending lots of hours to train those, and then it becomes kind of a IP or a way to use it. The secret sauce. Yeah.

Ferhat 19:11 Yeah. So they don't wanna share it. So, like, for example, you can think of this, like, for example, instead of the new car. So you put that model in in that in in in that car, and then it is or you can just point it, and then you can just step in and measure mass on it or power consumption because this is your car. Now we just wanna disclose it, so this could be kind of the use case for it. So the the other part, like, the our research was more deeply on focusing on a bit more. Okay. We can do some attacks on to reveal the hyperparameters of the models, like, obviously, how the delay is on that neural network. So how many nodes have we have on the neural network? What kind of neural modes they are using? So that was kind of our focus. But we were also trying to go deep on this, so to preview the, like, the exact parameters that is used on each of those layers. But it was a bit hard to attack in that specific model, uh, hardware for the HTP. But we successfully extracted the my colleagues were actually doing the most of the job. I was actually discussing the part of it, but, um, which is on this, um, mostly revealing the parameters of, like, what kind of nodes on that layer or what kind of, um, is it, like, convolutional network or fully connected nano network? Or is it, like, an add an additional layer or is it linear linear or nonlinear layer? So this kind of information we can extract for each layer, and then we can basically disclose each of those days during the late temporary attacking. So this is the kind of online, um, attacking attack so that each layer we can able to, like okay. We extracted this part. Okay. We know this is the, like, the fully connected layer with, um, 100 nodes on it. And then we can do the, like, the next next layer attack, and it goes until the, like, the end of the execution. So and then we can basically have better off idea of the model, and then maybe that could kind of give us an idea of the the, like, the the people who are actually trying to extract that information. So that was the kind of, yeah, important part that the company should take more, um, care on this to to design this kind of hardware.

Johannes 21:55 And then and all that is is is purely possible by by measuring the electromagnetic Correct. That are emitted by by the, uh, by the TPU and this example.

Ferhat 22:07 Yeah. I actually spend a lot of time, um, having that, that one item. And so, like, it was, like, kind of all 10 blood.

Johannes 22:16 Oh, you bet. Yeah. Yeah. Yeah. Yeah. Yeah. And and then with that information, were you able to, I don't know, is that the right term? Could you reconstruct that model in a way or? ah. I actually spend a lot of time, um, having that, that one item. And so, like, it was, like, kind of all 10 blood. Oh, you bet. Yeah. Yeah. Yeah. Yeah. Yeah. And and then with that information, were you able to, I don't know, is that the right term? Could you reconstruct that model in a way or?

Ferhat 22:28 Yeah, it's you have like the hyper parameters, so that is some information, but yeah, you have to have kind of, um, train the model again with the that inputs because, otherwise, you cannot have the whole model. But that basically improves a lot of steps like doing the training because in training, um, you are trying different models, different layers, and then, like, how many nodes. But it basically skip all those hyperparameter finding steps. And then you you have just, okay, this is the optimal model that they find it. Okay. We can train on this. So then we can just train, and we obtain the that specific parameters on them.

Johannes 23:13 And so, of course, you you you just said it. Right? I mean, AI models in in most cases contain, uh, a lot of either sensitive information or intellectual property and and data. So I think it'd be it'd be I'd be curious to hear on, you know, hardware level mitigation strategies, like, of course, there's photography, of course, there's encryption. Of course, as you mentioned, homomorphic encryption. There's obfuscation techniques. Um, so so what are the what what are what are the techniques that you feel are are best to protect against, um, you know, those leakych and tasks? Or is it a combination of all of them? ere's photography, of course, there's encryption. Of course, as you mentioned, homomorphic encryption. There's obfuscation techniques. Um, so so what are the what what are what are the techniques that you feel are are best to protect against, um, you know, those leakych and tasks? Or is it a combination of all of them?

Ferhat 23:52 Almost. Yeah. Um, that, like, three or four different way of doing it's to, like, do. So one is, like, kind of a masking. So, uh, uh, the masking is kind of, like, you are trying to put a a diverse kind of card to the lake the system. So whatever the system behaves, you're just doing the diverse part of it as in the masking part. So and then whatever the leakage is produced during that operation, it's basically diminished by, like, the other circuit. So this is kind of the part of it. Like, you can just put more, uh, design and hardware into it, and then, basically, the the the secret information that is revealed during the operation could be diminished, and it was the one. And the other one is, like, for example, hiding or shuffling. So the finding part is, like, it is not diminishing the leakage, but it makes harder for attackers to find specific combinations. So, like, for example, if you have a straightforward implementation in mind, But, uh, once you figure this out and then while trying to attack, you obtain some stuff, but at the end, whatever you obtain is not exactly is implemented. So so that was the kind of obfuscation you can produce on this. So that that prevent the attack from this. Yeah. That is the the domain kind of. And the other one is the, uh, maybe, like, introduce me some randomness onto system. So instead of hiding, so you can maybe put another system on through timing differences on the execution or, like, this introduce more kind of unknown part of the design so that basically behaves differently than normal. Then that could be the the way of attacking it. And then it could be diminished.

Johannes 26:12 And then and then a combination of those three, so so masking, shuffling, and introducing randomness probably uh, leads to the best results.

Ferhat 26:21 Yeah. That is definitely the like, um, it comes down to the decision of the late, what is the requirement for that design? Because you always, like, limit the design or the timing. So you have to give some kind of a decision based on that. So and that is the country that you have to receive, like, if you are trying to implement some some years.

Johannes 26:47 Yeah. Oh, and there's always a cost to it, right, as well.

Ferhat 26:50 Yeah.

Johannes 26:50 Yeah. You wanna, you wanna make something very secure, but very slow. Well then, yeah, you can certainly do that. Right?

Ferhat 26:58 But yes.

Johannes 26:59 I I mean, if we're talking about boot loaders, we're we're talking about milliseconds that, uh, that you wanna you wanna spend on this.

Ferhat 27:07 Yes. Yeah. Definitely. I don't wanna, like, work my computer. It takes more than a minute. Right? Or or my phone to opens up the light display or like, ] Johannes: I I mean, if we're talking about boot loaders, we're we're talking about milliseconds that, uh, that you wanna you wanna spend on this. Yes. Yeah. Definitely. I don't wanna, like, work my computer. It takes more than a minute. Right? Or or my phone to opens up the light display or like,

Johannes 27:17 uh, we, we just don't have the patience anymore. Right? Yeah. I remember, I remember the days when it was, you know, you switch on your computer and then you go get a coffee or you know, by the time you get back in. Yeah. Yes. So I think, um, I mean, we've we've we've been talking about, you know, um, homomorphic encryption a little bit. So Mhmm. I'd be I'd be curious to hear your your take on, you know, how to apply that to to machine learning and and, you know, protecting, uh, the AI privacy in in in AI workloads. And, of course, again, specifically, given that you, you know, work for a hardware vendor, specifically and you said it earlier, software is very different in in those hardware implementations. n, you know, how to apply that to to machine learning and and, you know, protecting, uh, the AI privacy in in in AI workloads. And, of course, again, specifically, given that you, you know, work for a hardware vendor, specifically and you said it earlier, software is very different in in those hardware implementations.

Ferhat 28:06 Yes. Correct. Yeah. The yeah. For the privacy in the area is, like, they still leave it on the phone because we always design that we have to see new everything efficiently and So these are the tools to to be considered on this action. So one of the things that kind of solves this problem, um, is the homomorphic encryption. So the animal's encryption basically does the same kind of processing in an in the encrypted environment. Even if you have the encrypted data, um, without decrypting it, you can get the regarding functional information you need. You can take that out of the that data, but you don't have to see everything on the data. So I'm not always kind of is going to be the the I can't actually say. That is, like, the most, um, essence of the, I mean, the whole machine function. So when they choose to make practical to do that evaluation and execution. So this people are actually working on this more than, like, fifteen years. So, um, because I get started with the, like, uh, 2009 PhD thesis from, like, gentry. Uh, he is a kind of the holy grail of on this, like, the finest one that could be still global, And then it started from home. So but, yeah, from all those fifteen years, um, we developed softwares, but it is still not actually practically doable. In order to to get accelerate or decide, we will ponder based on the the need for the. So usually one thing. So and even our applications have to be kind of revolved. We just have to, like, think differently and different differently from the homomorphic encryption. So, for example, one of my drives in Gandhi to the point, we are supposed to design genome analysis data. And then we were trying to classify the tumors based on the, uh, tumor type of based on the, like, the genome data of the patients. So we don't know anything about the, like, the genome data of that patient, but we were able to classify the, you know, the the type of the possible tumor type of that specific patient. So this was actually find a way to think in different thing because normally, we're not doing some kind of nonlinear operations, but normally in in linear operations can homomorphic encryption. You have to you can do this in the operations, you know, homomorphicly with that, like, spending or, like, introducing much much latency. So that was a thing maybe I've done previously on this. So I'm not currently planning. We are making more noise. Okay. We have this software stack or hardware stack to use AI. So how we can leverage that AI part on the lake on what we can coach you. So in in end, we were more focusing on this recent case. Right? We're using quantization, like, pruning methods, um, before the, like, the machine learning models. So can we also embrace these optimization introduced for AI in the home of the encryption space? So and then we we we observe that there is this determinedness increase if we can, um, design those optimizations for the homework encryption as well. So you can see some improvement on the execution, and then we can decrease the that wasn't it. these optimization introduced for AI in the home of the encryption space? So and then we we we observe that there is this determinedness increase if we can, um, design those optimizations for the homework encryption as well. So you can see some improvement on the execution, and then we can decrease the that wasn't it.

Johannes 33:40 I I I think it's it's fascinating to see, um, you know, with homomorphic encryption also, you you said the ideas, you know, more than 15 years old. And that just shows how, you know, how those timelines, um, you know, evolve in in in the cryptographic research and until until there is a practical use case and then, you know, technology still needs to catch up for us to to really reap the benefits of of things that are, you know, theoretically, mathematically possible and, you know, until until we have a a real world use case and and can apply those things and and frankly, environments that this can run on to find.

Ferhat 34:24 Yeah. Definitely. Um, for example, yeah, AMD is designing a lot of a 100, like a recent, uh, announcements on the, like, MI three fifty and, like, MI 400. So so those are the products that are accelerating. But in our case as well, so, like, can we leverage this information on on GPUs so that also uses homomorphic encryption to accelerate it? Or can we introduce some kind of a software stack or compiler stack that leverages optimizations for the that specific hardware so that we can also, uh, use that information and and apply to the the different hardware architectures. So that was the, uh, like the moving moving forward for for the homomorphic encryption space. You have to look into that those parts.

Johannes 35:20 For sure. So I guess, um, you know, pivoting a little bit back to to to side channel analysis. Um, of course, you know, there's tools to to support in in that work. So for instance, um, Keysight's side channel analysis suite, I think, has has been has been used and played a crucial role in in your research in particular. So so from your from what you said, how important are those, uh, you know, testing and validation tools for organizations that are developing quantum resistant? . So for instance, um, Keysight's side channel analysis suite, I think, has has been has been used and played a crucial role in in your research in particular. So so from your from what you said, how important are those, uh, you know, testing and validation tools for organizations that are developing quantum resistant?

Ferhat 35:54 Alright. Yeah. I mean, we have partnered with Keysight still for a long time. Like, I've been using their tools, um, like Keysight and Rescure. It was before Rescure. Um, so and I've been using their tools, uh, like, for, you know, academia before joining AMD, and then we were using we are using still. I have to use this side channel attacks. Um, there is also some open source attacks. And for example, um, sorry, open source tools. Yeah. Such as, like, Chip Besprit. So this is like another one that also, like, used, um, soft Egyptian attacks and such analysis as well. And things, for example, we are also partnering with Ansys to do, like, analysis in increasingly. So this was a reason, right, behind, um, we kind of leverage their tools to kind of design a new AES attack on classic crypto so that can be also, like, use this kind of attacks on the presale while using their tools? So, um, that was, like, methods of tools that we can leverage. But, yeah, there are still, um, a lot of different tools out there, and these are the things that we have evaluated and used so far. And, yeah, I mean, in in sectional space is, like, developing immensely. I wasn't expecting that much, but and then, like, a lot of people didn't know about, like, side channel, but, actually, um, these are, like, most on the physical side channel attacks. People mostly know about, like, spectrum and and more the, like, architectural side channel attacks. So those are the, like, the ones that actually have the lights and channel. And then it also, like, have introduced on this one. And then it becomes more recent too, and everyone actually trying. And it, like, open source tools and cheaper models. So this maybe means that, like, we can actually everyone actually can start doing these kind of things in the text. It not makes putting too much money on it. So we have to make sure that everything is protected in this in this world of, like, physical Slack channels and architectural Slack channels. So that is the the thing we have to take care about it, definitely.

Johannes 38:28 Yeah, and I think, uh, you know, your point on how front and center side channel attacks are nowadays, it's probably a combination of, uh, you know, in the past they were probably much more expensive, um, to conduct than they are now. And and then also, uh, you know, the the threat actors probably have changed as well. You know, ten, fifteen years ago, we weren't thinking too much about state actors and organized crimes to, you know, both have, like, very deep pockets. So that certainly plays a role in in, you know, proliferation of, uh, that being talked about as well.

Ferhat 39:08 Yes. Definitely. bably have changed as well. You know, ten, fifteen years ago, we weren't thinking too much about state actors and organized crimes to, you know, both have, like, very deep pockets. So that certainly plays a role in in, you know, proliferation of, uh, that being talked about as well. Yes. Definitely.

Johannes 39:10 So I think, you know, besides post quantum cryptography and homomorphic encryption that's, like, already very much, you know, forward looking. Is there is there any other area of innovation in in research direction that that you're excited about at the moment, especially, you know, thinking about post quantum photography and and, you know, protecting the privacy of AI models.

Ferhat 39:34 Yeah. These are the things that mostly interested for me laying testing for me. Um, Um, but, yeah, I mean, definitely anything in on AI is also, like, kind of always, like, spot on. So, like, whatever happens on the AI, you have to take care of it. And, like, so, like, it is so fast moving, so you cannot check off it. So and and definitely I'm always looking into, like, what kind of, uh, parts we can use that AI into the, like, the security space or sectional system. One of my work actually on this, uh, it's like on, um, my masterpieces about it. So I actually use, like, one of the recent, which are GPT to do some kind of engine on this one, like, sectional analysis. So I'm just kind of fine tune that part of GPT to, like, uh, behave as a, like, agent. And based on the, like, the outcome of our testing and, uh, like, the the models I I have trained it. So it was kind of giving a decision that, like, okay. You have to do, like, quite more transits to attack on this, or you have to try this different model to attack it. So it's kind of the way of attack. So this basically, um, creates more attack surfaces for attacker. And then that was also, um, there's also like the positive side of it that shows you that which you didn't think about it. So, like, the the way you you think, and that also, like, kind of countermeasures their proposal. But, yeah, that's still not that much powerful, but they're going there. So to, to use this like different agents to discuss about it, about the security and like come up with that leak more secure way of doing it. So that is always like kind of crazy, but yeah, it's it's

Johannes 41:46 it's it advances so, so quickly. Right? Every every day every day makes such a huge difference in in the way those models evolve. So, um, maybe you can give some some recommendations for, um, you know, organizations that are beginning to explore, uh, hardware post quantum implementations. Um, what are some, some concrete actions that you recommend to prioritize over, you know, the next twelve to twenty moments to get to a good security posture in that regard? you can give some some recommendations for, um, you know, organizations that are beginning to explore, uh, hardware post quantum implementations. Um, what are some, some concrete actions that you recommend to prioritize over, you know, the next twelve to twenty moments to get to a good security posture in that regard?

Ferhat 42:21 Yeah. For hardware security, yeah, it's definitely you have always the method with the final timing to give some kind of a decision. So so which quantum software, like, post quantum, I get it to have to, um, like, provide or, like, support is also kind of an important decision. And at some point, like, either or we will all going to use the post quantum cryptography. And the final phase is you cannot go back and change everything on the softwares as in software side of it. So you have to implement everything in the first day and then make sure everything is working correctly, and then there is no, like, no, should have gone in the first critical side channel or any other way of doing it. And maybe if we can have more to the client, like, if we can shift, like, the, like, post, uh, side channel security evaluations to the, like, piece like an error And then before, like, taking out the chip, so that would be, like, kind of the attack. Could be, like, more resilient to the side channels, at least. And then yeah. For example, in AMD, we are doing this, like, doing the the security analysis on on physical level that we are actually taking, um, uh, those actions or, like, side channel analysis before it could leak in the postcode, uh, postcode. And so that is always we are looking into it. And, yeah, as a hardware developer and like, they have to take into account on this one and then always listen to security guy on their shit.

Johannes 44:13 Yeah. He's he's your friend. He's not trying to slow things down. It's just like, yeah, security is not a bolt on. Even even more even more true in in the hardware business.

Ferhat 44:28 Yeah. That's great. Yeah. Definitely. Yeah. And sometimes, uh, we we we're, like, not precautious on this. Like, okay. We have to this in this day or, like, we implement this one or we have to take care of this. But, yeah, if we even like optimizations that we are think that they're not harmful, at some point it becomes harmful. So that's why you have to make sure that, that whatever your design have, you have to run those tests for the like side channel as well in the piece of it. So that makes sure everything is.

Johannes 45:10 Fascinating. Well, thanks so much, Farhad. Um, I think one, one last point, maybe, uh, I always like to ask where our listeners can follow your work or, you know, learn more about your research and get even more insights into hardware security, most fundamental photography, homomorphic encryption, AI privacy. ohannes: Fascinating. Well, thanks so much, Farhad. Um, I think one, one last point, maybe, uh, I always like to ask where our listeners can follow your work or, you know, learn more about your research and get even more insights into hardware security, most fundamental photography, homomorphic encryption, AI privacy.

Ferhat 45:30 Yeah. Sure. I yeah. You can look into my LinkedIn or, like, mostly, I have website, but I'm not updating it. I always think, like, writing a blog about it, like, for a more encryption and have to deploy the, like, last crypto and so and so forth. But I always post connect, Nate, and, actually, one of the fun things to do. But, yeah, if, like yeah. Probably, if I publication on my research, yeah, we are I'm still doing an AMD. So, uh, if you have published on this, probably, you will share on this, um, our, like, company website or or, like, given, like, four d website or or or.

Johannes 46:12 Absolutely. Well, thanks so much for joining today, Farhad. This has been, this has been very insightful. Thanks for sharing. Yeah. You know how.

Ferhat 46:20 Thank you very much for me, for inviting me. And that was very pleasure for me.

Johannes 46:25 Absolutely.