[00:00:09] >> Ok, Welcome to the i.s.p. lecture series. We are very pleased to have Jacob from Yale University to go by. Jim, I'm sorry I was mispronounced your name, but it's Zephyr. I see a writer that I thought Went to high school, partly in Poland, his origin country, and then Jacksonville, Illinois, University, Illinois, undergraduate. [00:00:42] I did a master's and Ph d., Princeton under our Ruby Lee who's well known in the Syrian security and hardware, software code design. He goes higher by e-mail, his assistant professor where he's been for, for the past 6 years and has published a book on one processor, a security Significant number of Grad students, undergraduate researchers. [00:01:11] James, It was nice not to invite me to give a talk last year Before the Yale Harvard football game which you know, won in double overtime. And I actually played it, you know, I was an undergrad, so I was very pleased to come visit Jacob. So I would, no further ADO will have Jacob take over, is going to talk about cloud security Jacob. [00:01:34] You can start to Take over a screen I think think think think you'll like f.p.g.a., cloud security and accelerated Infrastructures in the secure fashion. It preferable to think you can consent For the invitation and for the nice, nice introduction. So today I went back to get the part our work on securing f.p.g.a. accelerated our cloud infrastructures Island. [00:02:04] So this is basically at a work that we've been doing and continue to do I, since last Let's say to 2 or 2 years. Are we looking at the security issues associated with the f.p.g.a. which are used in the cloud and has been introduced me. And in Seattle and I'm leading the computer architecture and security lab i.d.l. universe. [00:02:25] So before I'll be Frank, get into the talk as well as the past. A little bit about our in our lab in the different types of projects we do. We actually cover a lot of a different aspect of architecture and hardware security. So today our focus on the of the, on the spur step, which is crowded today. [00:02:42] Security, which looks like, as you'll learn in the Park today, look at the f.p.g.a. that are used in the cloud. That is Amazon F One instances. And I just went single and looking tenant scenarios. And we're looking at basically different types of attack and how to security attack and how to mitigate them. [00:03:00] And especially we're not looking at just the security associate with the f.p.g.a. chip itself, but also the whole infrastructure. So the servers that are hosting their P.C.'s or the whole data centers are hosting the servers which are hosting that are the f.p.g.a.. And so that's what I'll talk about today. [00:03:16] But we also have projects related to hardware on especially on how to implement Ition are opposed quantum after progress in like with the progress piece. So these projects look at a, taking this novel path of a couple getting algorithms and implementing efficient hardware for them. So I'm, I don't claim to be a cripple grapher, but we try to do our best to create very efficient how do implementations again, focusing on f.p.g.a. after a prototype and has as uncensored and then Reduction. [00:03:48] My background is really from processor architecture security. And that's what I did my it says on and up. So when we continue to look at those topics, especially the different timing channel and processor caches, our translation of the sidebar 1st. And what doing work will now now kind of famous aspect and help the Apple attacks and a sort of the 4th force may major area. [00:04:12] We're looking at 100 security looking at different devices, for example, a physical collar will function. So how do you use and the physical properties of electronic devices for our, for security, such as for thinker printing or for Party generation and on it are to small advertisements. The one is the apparatus or security book that came out a year and a half. [00:04:35] 2 years ago now, so if you have access to the T.D.'s of the slides, all the links are click over to click the link to to see that I see that what books, web page and most universities actually you should be able to get a free p.d.f. to the Library, so it's you know, solid to spend money to, to read the book. [00:04:53] And the last advertisement as we're having so that you're an undergraduate looking pretty easy to supply to try Department of our group and also have a one or 2 a post doctoral up positions, especially in our, in our head. We're Crypto politicians and processor security. So the, if you're interested, you know, you can contact me or check out our, our Web site. [00:05:18] And then so tell US, I just want US to give you 11 idea about the overall research approach before I jump into the main talk. So I mentioned a lot of different topics that were, are we're exploring. And the reason we explore a lot of different topics is that they simply think as an attacker trying to compromise a system. [00:05:37] The only 2011 weakness in the system was the defenders. The need to defend the whole abhor system. You know this different tactic to that, that somebody might I come up with. So as defenders we really need to, we can't really focus on one topic, but we need to explore a lot of stuff. [00:05:54] So. So again, I was in doing a lot of work on c.p.u. and attacks and defenses, as are some recent publications to basically make sure that the c.p.u. is secured. But in addition to the c.p.u., the c.p.u. security attacker, Michael, after the accelerators are the f.p.g.a.. So this is what I'll talk about today, perhaps just a cute and this is kind of a, some publications recently we held at p.g.a. security. [00:06:18] And again, if you secure the f.p.g.a. Packer might go after things like the devices. So we also look at the camp a DRAM and, and other types of circuits to, for example, Bill, these are physically incredible functions. So you have to, as a security researcher, you have to consider all of those parts and art. [00:06:34] And there's also another one that, that I think gets much less attention is basically the infrastructure. So in addition to the, you know, the computer system or what's in your computer, a computer box is to have the whole, you know, and to structure of, of the data centers of power cooling. [00:06:50] I think like that we sort of look at, you know, the security of the, of the whole infrastructure of the computer system are I located as basically with security mind that you really need to defend all those others and kind of to help to help. We also publish a lot of open source code, especially hardware design, still up in the Hopper that gets people to use some of these defenses or to kind of understand the attacks that are right. [00:07:16] So we're jumping into a jumping into the main talk are not talking about the security of f.p.g.a. accelerator card infrastructure. This is really talking about the accelerators, f.p.g.a. part. And you know, if you have any questions, feel free to, you know, pack them in the chat and I believe Vince is able to see them and can, can read out the questions. [00:07:35] I'm happy to answer or our answer after that, after that Has a date and with a cover, some background about 2 program Gator Ace, or s.p. Chase and how they're used in the cloud. I'll talk about them, you know, information leaks and the **** infrastructure. So why are we worried about the security? [00:07:54] It's because of these as possible information leaks. And then I'll talk about Existing projects. And especially some, you know, at, towards the end about a recent ongoing work to get an idea of a kind of the, where the current state of the art is in terms of quality and security. [00:08:12] And then out of wrap up, wrap up the time. So what I call a car depreciates, I P T. Excellent infrastructure is what it is about. So basically there's this idea of cloud, public College. Yes. Where there's a card provider who you can see here has a co-sponsor of servers and inside it's server in a bunch of energy boards. [00:08:33] And then you know, they have many of these, of these servers with this huge importance. And then there are external card users who can basically on demand pay as you go, use these f.p.g.a. resources. So, you know, since 2000 And 120162017, when Allison Introduced the at the F one Excellent instances, you know, if you want to work with was higher with f.p.g.a., if you actually no longer have to have bad idea, P.C.'s, you can simply rent them are from Amazon or other cloud providers, Alabama Highway, all of the major providers have a p.g. offering. [00:09:11] And basically just as you are in temperature machine, you can rent a physical f.p.g.a. and you can work. You know, you can create hardware design and real hardware design without actually physically having to have that add into It. And so that the idea is that, you know, you can basically pay as you go quickly provision large number of, of f.p.g.a.. [00:09:30] And there's many, many providers that offer these services today. Other I'll talk about mostly about Apple cart. So Ok, so you have these f.p.g.a. to the cloud, you can, you know, it's a credit card to rent an f.p.g.a. and do whatever you want with it. But now, because Of this, people are interested and they're moving a lot of potentially step the computation to the cloud to be accelerated by the f.p.g.a.. [00:09:55] So this is a graphic, but at least for Amazon, you know, advertising that you can use and Genomics Research, financial cryptography, are you machine learning? These are the things that you can run algorithms on the f.p.g.a. hardware and an f.p.g.a. can post the performance and you can have passed the Pirates hellraisers part so far today they really lacked security. [00:10:16] And this is what kind of our research focuses on security of this activity. And, you know, why are we interested again, is there any information x. you could potentially intentionally think our financial data to program the keys? You know, maybe the machine learning model. So are there a lot of you know, can print F.p.g.a. from Amazon and run the sense that out of them. [00:10:36] But if the attacker is able to create some attack, they could take that information and lend, for example, your cryptographic keys or some, some information about that machine thing added up. So we want to understand what type of information exclude actually happen and what information you could get from these From these are appreciate and are the way to the information is leaking it through the, to the infrastructure other within either within the f.p.g.a. or across multiple different f.p.g.a., So again, you have pictures of this f.p.g.a. that a user rents in the cloud, but actually that is p.g. is not, not, not has that self is actually located in and in a server Next, other f.p.g.a. than the other servers that are hosting, you know, even more f.p.g.a. than that in the datacenter. [00:11:27] And then you know, so within one f.p.g.a. you can have some resources that are being shared. This would happen in a multi tenant car if p.j. standing when I want to put users are on the same f.p.g.a. or, you know, each user gets a separate f.p.g.a.. This would be the single tenant credit p.j. where everybody l.s.p. cheaper. [00:11:44] But even if you, if you rent or own, if it's a board program, Amazon, You still have some shared resources such as that. The p.c.i. E past are or the I or the other power power distribution, network power supply unit within the server is shared among all the p.g. boards cooling. [00:12:01] And then there's a lot of infrastructure that workings can they get the power that is shared with them of all these You know, these different servers and that p.g.a.. So that even though you're, you're getting an f.p.g.a. board that you're renting from a cloud provider, there's a lot of shared infrastructure. [00:12:17] And because of that sharing, there's potential for, or different information in the, or for ways to kind of compress, can transmit information. So to before talking about our some of our projects and if you talk a little bit about the threat model. So the main idea is that in kind of p.j.s., the p.g.a. itself is secure. [00:12:36] So we're not worried about physical attacks as an attack, I can rent and i.p.g. from Amazon. I can't physically go into the data Center and probe it. So a lot of invasive attacks or things like, you know, the m. ultimate medic radiation is not an issue because I detect it can't get into the data. [00:12:54] But the attacker can realize different potentially militia circuits on the f.p.g.a. board itself. And then use those circuits to try to learn or steal information. So they could, you know, they could try to create a career channel for leaking information. They could have a site channel to learn from another user. [00:13:12] To information from the shell to the shell is part of that that should be controlled by the card provider, but it's still implemented f.p.g. fabric. So you might be able to have shaken Yes. What's the d.c.t. This if you design a checkpoint, so with a sickly in, if you use it, you're familiar with with Xilinx that ailing store surveillance is one of the major f.p.g.a. providers in the credit c.g. board and the tools for Africa reading at a distance The hardware designs on the Xilinx boards. [00:13:49] And then when you create a 100 design, you usually create an edit on checkpoint, which is basically like you think you can think of that binary of your design and then US to what some, what the research proposal out there. For example, a car provider can get and the design checkpoint to see if there are any malicious triggers. [00:14:09] And then it'll, they could, you know, they could pipe some information and then, you know, actually I'll mention it later. For example, when you deploy your f.p.g.a. design on Amazon, they actually check if you have certain types of bring US later in the design. And they will prohibit you from are deploying a hardware design that has certain types of radar for later. [00:14:29] But are in, for example, in our own research to show that you can just design a different type of pre-nup flitter, that bypasses you know that there are checks that they're performing on the design checkpoint. So Yeah, but sorry, I should add that one at another Act, that's an acronym identified and also it is a fight of the a.s.i. is that Amazon f.p.g.a. image, which basically means that stream which is again the binary that you log on to the f.p.g.a. Port. [00:14:58] Cut and it's, so basically are, these are the attackers are basically asked able to, you know, create a, a militia circuit that they log onto the f.p.g.a. board. They can reconfigure the hardware at pretty much as much as they want. You know, you can, you can do some checking, but it might not catch everything. [00:15:15] And in a single Atena threat, basically the you can see my point to the attackers on one of p.g. board and then the other is the other user or the are the other attack. And in terms of the coverage channel on a different message board. So they might, you know, so you can try to see how information could make it across the board or in a multi multi country that should be a most attendant threat than a multi tenant scenario. [00:15:41] The attacker and the victim aren't as sharing the same at the same f.p.g.a. Port. And I can talk a little bit about how the shape happens later on, but to jump to, to kind of now talk about some of our research projects. So one of the, one of the 1st projects that we have looked at is how information needs to turn what channels basically to heat accumulated in a, in an f.p.g.a. chip. [00:16:10] Then add 2 sorts of ideas that you can basically raise the temperature on Earth of an f.p.g.a. chip and a different user later using that. If t.j. could monitor the temperature, so apt to how to do that. So one of the key things that used it's rather stuck is this idea of a ring oscillator, but this is that they stick a diagram of it's a ring oscillator is that is a sort of a loop of an odd number of, of inverters. [00:16:35] And then typically there may be like an end gate to enable it. But the idea that you have an input of a 0, then the 1st and vital change into one that has the Euro than a one. Then this one will pack the input with switch from, from a to your to a one. [00:16:50] And then again, the output of the 1012, basically an odd number of inverters creates this look that you forever trick late and off leaving one in a 0 and a free and free running. An oscillator, for example, can generate a lot of heat because basically the transistors are constantly switching and then on the other side. [00:17:10] And The frequency depends on the delay and the transistor switching which is affected by the Portage and the temperature. So the ring oscillator can be used both to generate a lot of heat off, but also at the same time it can be used as a sensor to monitor The temperature or the voltage inside the f.p.g.a. chip. [00:17:31] So for example, if you use, if you use the car provided doesn't give you a way to monitor the temperature of the city, you're in luck. I was the f.p.g.a.. They also do not give US answer that you can read the temperature. But if you can create a hardware design that has the ring oscillator, you can use that to, you know, very keyed muncher, the temperature without the car pride in knowing that you're actually munchie the temperature or the bolt inside that at the chip. [00:17:57] So if you're going to put up a bunch of these Aros together into an earthquake, can create, you know, 1010 thousands of, of ring oscillators that, that basically when you want to turn on, say they use up a lot of power, and they can significantly raise the temperature of an f.p.g.a. chip, and then on the other hand, you can create a r.-o. sensor where you basically have that one ring oscillator that, that, that oscillating. [00:18:23] And again, the oscillation will, the dependent on the temperature and the cottage. And then if you have a reference clock, you can compare the number of ring oscillator Lation to the number of, of the reference accounts, to basically get a relative idea, for example, about relative temperature with it getting higher or lower so. [00:18:43] So there are enough that it doesn't give you like a temperature unhealthiest, but the relative frequency you can basically check if the temperature is writing or far. So as to how to use this for information they get. So basically, what we created is, is a cover channel using the temperature of the f.p.g.a. Port. [00:19:01] And the idea is to use the thermal state of the f.p.g.a. Chip to send information between this the present call. If you want to send a binary one, you turn on a Big bring up a little array to heat up the chip. And if you turn and if you don't turn it on that the chip temperature stay, it stays constant. [00:19:18] So you transmit. You transmit a logical 0, and then since it actually takes an order a few minutes for the heat to dissipate, you can have a scenario where our Presented attack or use or rent one just as P.T.A.'s than RESI. They set up 3 of them while they do not stop one of them. [00:19:36] And then a, some time later, the user gives after a few j. and a different crowd of p.j. user comes in, asks Amazon to give them an f.p.g.a. and rent a friend, an f.p.g.a. from Amazon. And then they can basically create a sensor and monitor and from ations of it. [00:19:52] So after some time the user can recover, you know, the influence of the sender was trying to send 101111. And then after some time, because the boards are not fully called by the time the 2nd user uses it, they can read the temperature and recover then from age. [00:20:09] And this is just some kind of draft about the ring oscillator account. So basically that you can within a fixed amount of time, you can Count the Aro, oscillations, and you can fill in this graph. Basically, the board is heated up until time 0, and to at time 0 that the Big heaters are turned on. [00:20:26] And you can see the Arrow counts, and it goes from from a small number that jumped up significantly to the 1st jump. The Big jump is just a voltage change. The when you turn off at the heater a, the voltage At the voltage goes back up a little bit because the voltage regulators able to supply the, the correct about it. [00:20:46] So there's a Big jump in to the voltage. And then you can design in here is the Zoom in of the actual changes of the temperature. So after time 0, you can see that you have a few minutes before the temperature returns to the steady state. And just as the reference, this is sort of the that the temperature from a built in temperature sensor that will happen at p.g., but it will have. [00:21:08] But if you did this in Amazon, you don't get the top right graph because they don't project sensor to read temperature. But you can see that there is sort of relationship between the Arrow counts and the temperatures. You can use an Arrow to the monitor that the temperature And it's again on the order of, of it. [00:21:25] And basically One interesting thing is that the transmission channel is it's a sort of a non-symmetric about a channel that if you know the transmitter 0, that probably the option mission is 100 percent. Because if you don't see that p.g.a., it's not going to magically speed up. But if you 10, if you try to transmit a one, there is some probability that the, this one is going to flip to 0 because that p.g. could qualify between the time he set up and when you and receive that information. [00:21:54] So those kind of a non-symmetric communication channel, but you can account for that with correct correct code. And since we are able to transmit one bit press p.g.a., you can go to Amazon and rent hundreds of the g. 8. If you have a Big enough credit card, so you can scale the channel bytes, you can see, you know, if you use one f.p.g.a. versus renting a 256 f.p.g.a.. [00:22:20] So we see this kind of, we presented the 1st thermal importunate P G 8. And we actually demonstrated in the specs of that and Computing Center which is over different from Amazon. But Yes, the Miller had, you read f.p.g.a. and different users can have access to differentiate and, you know, put them by renting or f.p.g.a. to can increase that. [00:22:40] And with their many research opportunities, for example, one would be increasing the bandwidth or trying to transmit multiple bits for the project. So it's out of the 1st step, but they're still really interesting research how to improve Ok, so I'll jump to jump to next question. I'm trying to get to 3 projects here, so if it is a little past, but I'll show that slide if you want to review them later on. [00:23:08] So the, the 2nd project is on information, the 2, I shared power supply. So in the 1st case, we are basically looking at the temperature so we can take up the cooling infrastructure is the thing that is being shared and then the 2nd in, in this part, I'll talk about the shared power supply. [00:23:24] But in this case, the power supply is what is shared among different And so Basically we kind of wanted to, can we transmit information through a shared resource, such as that a shared power supply. And Here's a diagram of a basically of what happened in, in one of those a car. [00:23:45] If you see a perverse attitude happens in your, in your own workstation, or in your computer where you have a power supply unit that basically provides a 12 volt input to all the devices in the server. For your workstation that kind of you might be familiar with, like the $610.00, MOLEX connectors are 12 hours and then the 12 volt goes to the, you know, to the g.p.u. to the c.p.u. or to the f.p.g.a. board. [00:24:10] And you can see that, you know, although that all these components are separate, these Red lines show that there is a connection via the 12 volt lines between the g.p.u. and that p.g.a. or the most oppressed p.g.a. source if you left it at. So what happens is that there's this shift power supply, and an ascending or a transmitter can modulate the power consumption. [00:24:33] On the t. on the f.p.g.a., you can again create a huge circuit that drives a lot of power. And if you draw enough power, that the power supply is actually not going to be able to supply, you know, they'll be a small dip in the 12 volt to ply. [00:24:46] And then what you can do is you can actually have a different f.p.g.a., which is able to sense the change than those voltage to get without. By using oscillators which are sensitive to both exchanges. You can actually tell some small changes on the voltage. And so you can have a tender transmit information to another f.p.g.a., or actually you can abstract with between C.P.U.'s the p.g.a. can or g.p.u. an f.p.g.a. where again, if the c.p.u. or our g.p.u. is a lot of computation, that would drive a lot of power bring down to the voltage a little bit, but just enough that the arrows on the f.p.g.a. will too to not And, and so this is sort of a diagram of this, of this attack where you can have the source at the G 8 which is transmitting the stink at the G 8, which is receiving and look at what similar to the thermal channel again, you know, you keep seeing these remarks later is again that on the source f.p.g.a.. [00:25:44] Here we use a large array of ring US later to consume a lot of power and stress, the power supplies. Of course, they'll be a thermal effect that talked before, but there's also an additional effect that you can stress to the shared power supply. And then to the share power supply between the 22 point and On the receiver, you can see to 2 components of this that the census of the Blue ring oscillators are the sensors, which again the oscillations of the ring oscillator depend on the voltage a little bit but you can detect all the changes by, by checking that up the ring are that the ring oscillation are changes. [00:26:23] But also we have this idea of a stressor are, so I can talk about maybe later questions. But basically you have this 12 volt wattage coming into the f.p.g.a. and I are left e.g.. There is a voltage regulator that brings the down to a one volt. But if you have some small changes in the 12 volt input, the onboard regulator is still able to kind of correct that and give you sort of constant one Port input to the f.p.g.a.. [00:26:49] And this one pulse is really what there are old are measuring. But so what would the solo for that sort of mask and changes that have in the 12 volt at share between the different f.p.g.a. at, by introducing the stress and I wrote, were actually able to stress this 12 to one halt wattage regulator. [00:27:09] So that if I get it stressed enough, I think this is an app power drawn on the local f.p.g.a.. Again, if there's a small change in the 12 of input, it's not actually able to maintain that one volt output. And you can observe that the sort of like a 2 step at a step process. [00:27:25] And Here you can, you can see some graph, I think the 2nd one who will explain the stressors quite, quite nicely. But the 1st graph basically kind of intuitive, you know, if I have more transmitters, if I have, if I draw, draw more power than the transmission, the accuracy keeps increasing. [00:27:44] Not in some cases on 100 percent, so we are so k.c. and AC. So cases that complex are 7 series ports from Xilinx, an AC Ide Arctics are 7 series. But from my balance its interest Is a quick chat question. Do you have noise issues from other refugees? [00:28:04] Possibly Exact perfect Quest? Yes. Yes. So you can have exactly noise issues from, from other f.p.g.a.. Or if I could actually go back a little, but you can have noise issues even from the city or the g.p.u.. So in the, in the control settings that, that we could experiment in. [00:28:25] It was basically just, you know, the sender and receiver. So we didn't, didn't create intentional noise, I.D.'s. And If Yes, there's a lot of activity on different f.p.g.a. then that would, that would include introduce noise issues. So you have to Account for that exactly. What was, do you have? [00:28:48] This is called knowledge of where embers are and gives you the f.p.g.a. simmered. You have identification of location provided And when not to Amazon does not provide that information. We. We have reversed engineered enough to get some good ideas. I will, I will not exactly talk about this in this talk of the sort of still ongoing. [00:29:14] But we were, were basically able to collocate up to S.P.'s on the same server in the Amazon without any public information. They say On different types of information though. Ok, but I'm not going to grab it. And I just want to go back to the power question this really quickly. [00:29:36] On one hand you can have noise, which is sort of negative, but that can also be a site channel, right? So if somebody is doing a lot of g.p.u. computing Ition, I can be running my sensor on the f.p.g.a.. And I'm actually possibly learning some information and you know, when there's cheap You activity or not so hard once in the noise can be a bad thing. [00:29:56] If you would, if were trying to covertly communicate and somebody is interfering with US. But on the other hand, if it's a, it's Information Act that, you know, if you just have this I channel, you can learn about this if you activity or that if you activity From, you know, by monitoring the power so that, that, that, and so, and into this I want to so that next stop was talked about the accuracy 1st of the number of stress risk. [00:30:23] So remember on the, on the receiver there are the stressors. And this is very intuitive. But it's kind of comes out nicely in the graph, is that as you have more stressors, It stresses the law of the trade later and your and that helps you observe the external, The 12 volt changes a bit better. [00:30:40] But then at some point, if you stress it too much, then the local, you know, basically dropped the ball to so much that, that sort of masks any, any changes to the external 12 or changes. So when you're doing that, the sensing have to kind of balance just in Astra serous, to get a good accuracy, but not too much to, you know, just this keep create South, create a bunch of noise and cancel any and transmission chances. [00:31:06] So that's kind of an interesting, an interesting point that happens a couple of New stressors. And then finally, Also occur pretty intuitive. You know, as you really need to increase more measure, you know, repeat that, the transition to prick over a channel. If you keep pretending the same information 200 times then still reach honey, 100 percent accuracy, And yet such as in just kind of sums coming in. [00:31:34] To summarize some results. Note we tried different different transmitters with different receivers. The c.s.u. different types of up power supply units are inherent to 2 different types. And you know that channel accuracy with a few 100 repetitions is It's quite reaching 100 percent and it does vary by the power supply units are present. [00:31:57] This is legit to you to f.p.g.a. channel. But more and more importantly can see that different p.s. p.s. use had different different accuracy. So while we know we don't recant often about what type of power supply units in our server. But I think this work shows that, you know, Let's say basically paying for a Forum where Phantom power supply that has, you know, maybe more, more, more filtering or more As high higher rating would, will be able to mask some of these channels. [00:32:30] Which is, you know, a commodity thing that has a p.s.u., it actually becomes an important factor when you consider, consider the security and the interesting thing. And I already talked about that in the lab that we have to few to f.p.g.a. and g.p.u. To f.p.g.a. channels. And we can restore our a bunch of our research opportunities for Crap that other people here are interested to look at. [00:32:57] I think one of the things I will, I will not talk about, but in addition to ring oscillators, there is this hand to digital converter, which is the Captain at the top of a circuit that has A very clear the delay in the time to digital converter. Has a very clear relationship between voltage and, and the delay of the circuit. [00:33:19] So you can use to see it as a voltage sensor. And actually our collaborators do math and use them quite successfully. So using T.C.'s for these power channels is another, another opportunity and they should actually, if you even better results because there isn't more sensitive to that to or to change Cook. [00:33:39] All right, so now again another, another, a project I want to talk about, which actually gets, gets a little bit to the question that that was posted about, you know, do we have some idea about where things are, are located in the data Center. So we know the finger printing does not exactly let you do that, but it does let you actually identify unique. [00:34:04] I Speech a board. And more importantly, identify if you are currently using an f.p.g.a. board that you have used before because it has a thing. Think again that the card providers are to, To protect their infrastructure. Don't give you any information about any sort of like a serial number. [00:34:26] If you can't get a serial number of the f.p.g.a. you are using in Amazon. But by creating add pressure circuits for fingerprinting, you can effectively get a serial number for each f.p.g.a. and you can identify so out and if you know it and it's not just, it's not just a tax, but it can also help security. [00:34:44] Or for example, when the, when you know, when that had provider promises you that, you know, that each time I rent an f.p.g.a., you're going to get a different f.p.g.a. to make sure you know there's some reliable jury Dundon. See, you could use the think or Prince to actually verify that Yes, I've gotten Air at different f.p.g.a. the standard I got last time, as opposed to, you know, doing a, in a fantasy message each time. [00:35:08] So it's not to all of these are, you know, they're not, not just attacks But you can use it also for defense like for getting up.