Assistente AI
Trascrizione
00:00:30Alessandro Brighente: There we go.
00:09:210Alessandro Brighente: Oh, finally, we'll
00:12:540Alessandro Brighente: good
00:14:300Alessandro Brighente: everyone. So
00:16:490Alessandro Brighente: let's get back to the
00:18:940Alessandro Brighente: competitive control. Right? So last time we saw how we can model this because multiple vehicles and all of them needs to have this
00:31:460Alessandro Brighente: common ground or control strategy right? And not only we can model single vehicles, but each vehicle can design a model for the whole platoon to account for. For all the other vehicles
00:47:600Alessandro Brighente: something that I didn't mention assignment I realized was a bit confused. Is this thing about the the design headway? Right? So what is the headway? If you if you do the math. You see that we are having a distance in here.
01:04:646Alessandro Brighente: This is speed distance again. So this needs to be time. Right? So the the constant headway that we use for for vehicles and control strategies in vehicles is time. So the the this headway tells us how much time we want to separate
01:22:708Alessandro Brighente: to vehicles, right? So if they're driving at a given speed we want these headway to tell us, okay, how much time will they require for the vehicle to crash into the morning from right? And this gives us, decent for
01:41:90Alessandro Brighente: sufficient amount of space. Right? And yeah, here you have all the equations for this.
01:48:242Alessandro Brighente: But the last thing that we were discussing
01:52:07Alessandro Brighente: where the attack strategies, right? So
01:55:663Alessandro Brighente: which kind of different attacks can an attacker exploit in order to cause misbehaviors, or, create some anomalies in the platoon, making the 4 biggest crash.
02:09:820Alessandro Brighente: Good. So the the problem is, how do we detect these attacks? Right? How do we defend against the fact that even though we might have a cryptographic means protecting the communications still, part of the part of the platoon might be malicious, and therefore deliver a valid messages with the malicious content in there.
02:33:650Alessandro Brighente: Good. So again, we have our control strategies. Right? This is the double integrator model. It's its name, basically, tells us the the control strategy of the platoon right? How do we want it to behave this small dot on top of the letters, not the derivative derivative in time. And indeed, you see, the derivative of speeds. Acceleration of space is
03:03:974Alessandro Brighente: the the speed. Okay? But yeah, we don't really care about the derivation of this control strategy. We just know that this is the the way the the vehicle behave right. And you see the error in here, right? The error, the error is given by this difference of speed between I and we called I minus one.
03:29:170Alessandro Brighente: and the acceleration of the vehicle great. So a lot of things to say we want to minimize the error right where the error is keeping vehicles at a given space right? We don't want the space to do to decrease too much or be too big, because otherwise, if vehicles are too far apart they might lose connection, and we might disrupt the tool right. But we want to keep the distance between
03:56:272Alessandro Brighente: couples of people in the platoon
04:00:270Alessandro Brighente: at the given target value.
04:03:330Alessandro Brighente: Good. So
04:06:970Alessandro Brighente: as we did before we define the state of the car. Right? The state of the I car is given by the error value, the speed of the acceleration and the the the control value coming from the other vehicle right from the vehicle in the front.
04:24:260Alessandro Brighente: and the way in which we can write the state of the integration for for the Id card is through linear systems. Right? We have this linear system in here. Which, what this thing is basically doing is computing a set of linear equations. Right? If you substitute these
04:44:320Alessandro Brighente: A I in here. Through in these equations, you will basically see that the derivative of this Xi is just a linear combination of the Xi Xi, minus one ui and ui minus one, right? So basically the the the location in space of the I vehicle of the I minus one vehicle
05:09:988Alessandro Brighente: and the control signal of the I vehicle and the estimate of the control signal from the vehicle in front. Right? So the thing that we are receiving through through communications.
05:20:850Alessandro Brighente: Okay? So this one said, describes the state of the very 1st vehicle. Right? The the 1st vehicle doesn't communicate with other people. It doesn't receive control signals from other vehicles, because, indeed, is the the 1st one. You assume that that one is the leader. So it's just sending information to the other vehicles. Right? It's not receiving anything. And that's why in here you don't have this term right? There's no
05:43:760Alessandro Brighente: minus one, Nico, and we don't have this one in here, right? Because basically, we are not receiving anything. Okay?
05:54:30Alessandro Brighente: and yeah, so these linear equations are just implemented. The double integrated model that we've seen before.
06:01:440Alessandro Brighente: Okay? But again, the important part in here is that we can model the behavior of the vehicles. Right? We can design. No, we can.
06:13:354Alessandro Brighente: yeah, we can design a prediction model, right? We know.
06:18:20Alessandro Brighente: what
06:18:980Alessandro Brighente: we expect the vehicle to do in a successive time instant. Right? It means that if I know that I'm driving at the constant speed, right? I'm driving at 50 kilometers per hour and I'm doing that the time? T. 0. What will my state be? A time 10? Right? So I just know that I'm going in a straight line after 10 seconds I would be
06:44:60Alessandro Brighente: at a certain location. This is exactly what this thing is. Here is the strategy, and we are also accounting for the fact that we should keep the distance between successive people as we need to control it right? We want it to be close to the target and the fact that we are receiving some information
07:03:380Alessandro Brighente: from the other vehicle. Right? So if the vehicle in front is going to to stop, and that's something that we need to take into account. Right? So these scenarios are just some in the effects of my behavior. The the behavior of the vehicle in the front and the the information that they got through through the control strategies. Right? The the uis.
07:27:930Alessandro Brighente: Okay. So given that we can do this for a single vehicle. It means that we can do that for the whole platoon right? If we have the description of the I vehicle, nothing prevents us from having the description of the I minus one or the I plus one.
07:43:80Alessandro Brighente: Right? So we can have a very big system of linear equations that describes the whole platoon. Right? Indeed. In here. You see, we have this capital X in here. Which is just collecting the the the x values, the location space of all the vehicles. Right? We assume to have
08:02:210Alessandro Brighente: capital K vehicles. We define the inputs of the systems. Again, as is vector in here, the input of the systems of all the vehicles in the car and the in the platoon.
08:16:140Alessandro Brighente: And we have the equation for the whole system right again. So what we have is that the derivative in time or the X values are again a linear combination of all this in the
08:30:730Alessandro Brighente: the the states of the cars in the in the system. I do. I mean.
08:37:380Alessandro Brighente: it's daily
08:39:140Alessandro Brighente: straightforward, where you can just combine the the whole
08:42:713Alessandro Brighente: the whole set of equations that we've seen before, and put them in here. You will see that you have quiz
08:48:680Alessandro Brighente: these very big margins with the the whole set of variables.
08:56:630Alessandro Brighente: Of course, we cannot work with the derivatives. Right? I mean, we can do the math with derivatives. But then the system needs to. We need to write an algorithm. So instead of considering the the system as
09:13:390Alessandro Brighente: in in the continuous time.
09:17:60Alessandro Brighente: we have a discrete version of the system. Right? So instead of writing the derivative, we just consider time here. Right? So we said that the X
09:31:531Alessandro Brighente: the X value a time k plus one.
09:35:920Alessandro Brighente: So we replace the the.in here that we had in here before with this k plus one or comfortable type that we are predicting. Somehow the future is given by the cons, the the current value of X
09:50:130Alessandro Brighente: and the current value of the input to the system, right
09:56:440Alessandro Brighente: and
09:57:850Alessandro Brighente: this thing here defines our control strategy. For so for the I vehicle, the control strategy is given by again a linear combination of the state of the vehicle at the time. K, and the state of the same vehicle. Okay? So again, we don't do the math of why, we have this specific control structure. Why, that equation provides us with certain properties for the system.
10:24:40Alessandro Brighente: But we just care about the fact that we know that the the state of the system, right? Or the way the system behaves is just a linear combination of different effects that people
10:37:670Alessandro Brighente: goes through right? The different effects mean. It's a state in space station space the way it behaves and the inputs that it gets from from the other vehicles.
10:53:550Alessandro Brighente: Why do we care about all these things? Right? So we said that our problem is the fact that some of the vehicles in the platoon might be malicious and send us malicious instructions. Right? So, for instance, the
11:09:803Alessandro Brighente: the people in front of me sends me a comment telling me to to stop right or to to speed up or doing something that the other members of the platoon are not doing right. And we want to to understand whether this comment makes sense with respect to the whole state of the platoon. Right?
11:28:780Alessandro Brighente: Good. So how do we do? We do this? Well, if we have means to model, the behavior of the the system of the all the cars that belong to the platoon, the composite platoon.
11:41:299Alessandro Brighente: Then we have means to verify whether the information that we are receiving makes sense with respect to the evolution of the system. Right? So it's something like, I know what to expect from from my system, and I need to verify whether, what other vehicles are telling me makes sense with my expectations on how the system will behave in time right? And we have that thing. We have that means of
12:09:140Alessandro Brighente: Having these expectations right? They're predicting
12:12:30Alessandro Brighente: how the the state of the system will evolve because we have these integrations in here. Right? All these things are just telling us what we expect the I to do at the certain dimension.
12:25:540Alessandro Brighente: and remember that this
12:28:350Alessandro Brighente: equation here is telling us the state of the whole platoon. Right? So at each car in the platoon, I can have this set of equations, right?
12:39:470Alessandro Brighente: Thanks to which
12:42:860Alessandro Brighente: each car will have its own version
12:47:610Alessandro Brighente: of predictions for the the whole platform. Okay?
12:51:760Alessandro Brighente: So they can cross check the the different information that they have. It's not like we have a centralized controller with a single model in there. With a central point of failure. Now all the cars will have their own version of this prediction system. Thanks to which they can check whether the instructions they they receive make sense. With respect to these predictions.
13:13:950Alessandro Brighente: Okay, so here's a block diagram of how the model based detection would work.
13:22:200Alessandro Brighente: Right? So we have the the information that the vehicle itself has right so the the estimate of the distance with the vehicle in front the speed of the vehicle in front its own speed. It then calculates the address right? And try to
13:47:780Alessandro Brighente: to see whether they receive instruction match with the the expectation system, and then see whether there's an error or not. Right? When do we have an error? Well, we have an error. When, the instructions that we receive from the other vehicles in the platoon do not match with our prediction.
14:08:710Alessandro Brighente: of the the behavior of the system right? And when we detect an error, what do we do? Well, we can choose whether to stick with the with Cacc policy, or whether to move back to Acc and do not consider anymore the inputs that we get from from the other vehicles.
14:31:110Alessandro Brighente: Okay, so
14:34:245Alessandro Brighente: how does this thing work
14:39:80Alessandro Brighente: again.
14:41:70Alessandro Brighente: So to move a bit down.
14:48:440Alessandro Brighente: Good. So how does this thing work? Well, we said that. thanks to these Ccc
14:54:750Alessandro Brighente: we have 2 sources of information. Right? The 1st one is the value that I can measure with respect to the vehicle in the phone, right? So I can compute the distance that separates me and the vehicle in the front, and then I receive information from the other vehicles through through communication means right. So I receive some data packets through the Slc. Whatever that tells me. Some information about the vehicles.
15:21:390Alessandro Brighente: Good. So let's say that again, we consider
15:26:00Alessandro Brighente: the I've car.
15:28:250Alessandro Brighente: and we want the Icard to to be able to model the behavior of the I minus one
15:34:840Alessandro Brighente: car given that, it receives the package from a car, I and say, right? So we have some of the information of the car in the front and some other car that is farther away in the blue.
15:49:718Alessandro Brighente: Good. So we can define this model. Right? So when we have these emphatics in here. In the notes, the model that we create inside the the the control module of each car. And we just
16:04:80Alessandro Brighente: seen some slides ago how we can model the behavior of the platoon. Right? So our model is exactly what we've seen before. Right? Selinar combination on the status of the current vehicle. Locations. The time K and the input signals that we get from the the controller and the other vehicles. Communications. Again at times. K,
16:30:560Alessandro Brighente: okay, so we we do use exactly the model that we've seen before
16:34:750Alessandro Brighente: to have something simple.
16:38:650Alessandro Brighente: Good. So we also know, the control integration right? The control strategy that we use before. We had the question mark with the one some slides ago. Right? This thing here, this is the control strategy. Right? So given the States.
16:55:380Alessandro Brighente: Whatever target we are considering. This is how we're going to build the the value you, the control strategy, right? The control input that we go, that we give to the to the car.
17:06:180Alessandro Brighente: Right? So it means that. If we receive some information from other cars. Right?
17:19:109Alessandro Brighente: let's put it this way. We have,
17:22:490Alessandro Brighente: 2 different settings in in the platoon, right? We have times in which we actually receive packets from other vehicles and times in which we don't receive packets from other vehicles. But still we need to update the our control strategy, right?
17:37:300Alessandro Brighente: So we can see these as as lots in time. At certain slots, we get information from other cars. And in other slots, we do not get such information through communication. Okay, so this is what happens when we get the information from other other cars. Right? So we talk about the update period. And this is our control strategy through in these update periods. Right? Again.
18:02:800Alessandro Brighente: a linear combination in the state of the of the card. And this information that we get through through communication. Right? So what we get from the I minus J card. What the that card tells us is actually doing.
18:22:630Alessandro Brighente: Instead, when we have a known update period. Again, the control strategy would be the same. But we know we cannot rely on the information that we get from from the other card. Right? So we do not account for for that term in our control strategy. Okay? So the only difference is whether we include that information on on our update patient or not.
18:58:780Alessandro Brighente: Okay, yeah, this is just telling you, okay, we have a simple model. Right? We just have linear equation. We're just combining linearly combining the States of all the vehicles in the platoon. And we are performing some prediction on the
19:15:194Alessandro Brighente: I and J. The vehicle right? Because we have a model for all of them, so we can predict what we expect that vehicle to tell us at a given moment in time. Right? So
19:27:950Alessandro Brighente: if that is the case, it means that we can create models that are as complicated as we want. Right? We? This is one example, right? It can be more complicated. But we actually don't need that. So the problem is, how do we integrate this thing about the computing errors, right? Finding mismatches between our predictions and the communications that we get from
19:50:605Alessandro Brighente: from the other people. Well, we have our model right? Our linear model. So
19:59:20Alessandro Brighente: out from our linear model, we will have a specific value and specific value of what we expect a certain vehicle to tell us, and we compare it with what it actually tells us. Right? So you might have discrepancies in there, due to the fact that your prediction model might be. Not that good you might have discrepancies due to
20:22:264Alessandro Brighente: some random event occurring, or some noises on top of the control signal right? So we need to account for that. We cannot expect to have always 0 values so what we do in this case is to set the threshold value. We say that we have an error. When the difference between our prediction and the value that we receive from the car is above a given threshold. Okay?
20:44:855Alessandro Brighente: And that indicates that okay, whatever might be happening that might deserve our system. This is definitely not the a valid input, that we get from from that card.
20:58:440Alessandro Brighente: Okay? So we need to define some some models. Okay, so here you have an example on how you can. you can compute errors on different values. Right? You have the add on acceleration, on the speed and on the the value that you get from
21:14:727Alessandro Brighente: through communications. Right? And these are all the errors that you can compute based on the U value that you get from the other vehicle. Okay? So given that specific U value, you know the the control strategy of the vehicle. You just can
21:31:75Alessandro Brighente: put this value in your state update equation, perform your prediction, and therefore have predictions on the expected acceleration of that people on selected speed of that people and the expected control strategy of the people. Given all these things in here right? And then you want to set threshold on all of these values in here. Right? So do they make sense with respect to to our model or not.
21:58:107Alessandro Brighente: Good. So here you have, some of the results. Based on a very simple system. Right? So the the idea here is that from a vehicle in the front we receive a message telling us to so
22:13:260Alessandro Brighente: right, and that would mean
22:18:870Alessandro Brighente: a decreased distance between the vehicles, right? And that's something you do not want.
22:23:714Alessandro Brighente: So based on your sense values and your prediction model, you can decide whether this top signal is valid or not. And if you do not use any mitigation strategy, this is what happens right? You have a distance in here at a certain point gets to 0. It means that you crashed up with the with the legal in the front. Instead, if you have the mitigation strategy. You see that you get close to 0. I promise this is
22:49:360Alessandro Brighente: very close to 0. But it's not 0. Okay, so you're not not actually crashing. And then you get to a single state as well what is happening here? We are comparing the the model, our predictions with what actually happens right, the distance with the vehicle in the front, and the information that we receive from the other vehicle. So at a certain point, we realize that something wrong is happening here. And we just react to keep the distance
23:15:450Alessandro Brighente: at the the reference value.
23:21:90Alessandro Brighente: Okay, so
23:23:580Alessandro Brighente: this seems to be kind of complicated. But it's
23:27:760Alessandro Brighente: it's really not in the sense that all of these models are something that are very well known. So if you if you check, for instance, you have implementations of people, platoons in many different software.
23:43:960Alessandro Brighente: see you.
23:46:70Alessandro Brighente: So, for instance, you have examples on how to model platoons of trucks in math lab. You have them in python. You have them in color. You have them in Ross. You have them on a lot of different nuances right? And you don't do not actually care about the control strategy because it's not your job to design the control strategy your job would be okay, I know that this is how I'm modeling my vehicle. Right? This is how the different vehicles respond to different inputs that they get
24:14:470Alessandro Brighente: okay. And how do I create a model that tells me where there's an anomaly? Right? The concept of anomaly here is very simple is the diverges between what we expect the people to do, and what information we actually get from the environment or the other people's right. And it's just a matter of checking the mismatches. This is one of the things that you will have for
24:40:670Alessandro Brighente: So for the mid course project you will have an example already implemented on how the platoon behaves right now, how the controls behave, and you can check whether you can detect this kind of of anomalies or not
24:57:610Alessandro Brighente: So of course, here we are talking about something that is that goes on the term of model predictive control. Right? We have performance or some predictions we know, the behavior of the model. We know, how we expect the the system to change in time.
25:14:588Alessandro Brighente: And we can again make it complicated as much as we want. So we can, for instance, model just the vehicle in front of us. We can model the whole platoon. We can model what we want. So this is one approach. But another approach can can be slightly different and based on
25:35:938Alessandro Brighente: on the communication means right. The communication means what it means We would like to somehow know whether the information that we received has actually been generated by just a member of the platoon.
25:50:300Alessandro Brighente: and usually what we have is in this scenario that we we trust the platoon leader. Right? So we have, an entity that is, secure. We assume it's secure right, and it's the head of the platoon usually. So it's the one that is conveying control information to all the other members that need to follow along the leader.
26:09:840Alessandro Brighente: Okay? And so what we assume here is a replay attack. Right? So it means that at a certain point the platoon leader sent a packet to one of the the platoon members. An attacker captured this packet right and the attacker just replace that later in time.
26:31:140Alessandro Brighente: So again, what is the fact of this attack? Well, the the packet, whatever information that the attacker is conveying to the the victim is valid right? Because it's a packet that has been generated by. By the by, the
26:45:900Alessandro Brighente: so we'll be
26:47:230Alessandro Brighente: deemed as valid as control rights. For instance, we'll have a valid signature will be authenticated in the sense. Right? So how do we detect these these kind of attacks without
26:58:870Alessandro Brighente: changing the protocols that we have in these vehicle to vehicle communications
27:11:164Alessandro Brighente: so one approach to detect these repay attacks is to hide some
27:17:620Alessandro Brighente: signature, hide some secret information in the in instructions that the platoons leader sends to the platoon members. Okay? So here we have the way in which we can change the control signal that the platoon leader sends to the other vehicles. Right? So the idea is
27:38:440Alessandro Brighente: This. Ak, 0 in here is the control signal that the platform leader thinks is valid. For this, I mean surprising, just sends it to the other vehicles in the in the platform.
27:50:173Alessandro Brighente: The idea is to embed the a noisy signal in here right? Some random sequence
27:56:60Alessandro Brighente: that only the the platoon leader can generate, and that it can send to the platoon members such that they can verify the presence of the signal in their control strategy. Right? So the idea is that thanks to these noisy value here.
28:16:473Alessandro Brighente: All the platform members will update their safety issues
28:20:870Alessandro Brighente: embedding this noise sequence. Right? So so let's say I'm the 1st vehicle after the leader. Right? I will receive the control signal from the leader right? And it will embed this noise in my control strategy. Then I will forward my control strategy back the vehicle
28:40:870Alessandro Brighente: on my back right. But my control strategy will already embed the noisy signal. So whatever is sent to the 3rd vehicle will embed the noisy signal as well, which performs their computation, and then send it to the other people in the back again and basically propagate these this noisy signal in all the communications that we have.
29:06:330Alessandro Brighente: Okay? So also, in this case. In order to to check for the presence of this noisy signal, we need the a virtual model of the platoon right again as we did before. What we would like to do is to estimate the State in the future of the vehicle in front of me, for instance in the absence of every play attack? Right? So if everything goes well, how do I expect the acceleration of the vehicle in front of me to be
29:35:412Alessandro Brighente: in the the next future. Right? So
29:40:360Alessandro Brighente: we use something that is very similar to what we had before. Right? So again, we have our state of the equation, which is just a linear combination of the state of the the system.
29:53:370Alessandro Brighente: A time k plus the noisy signal right? We want to account for this noise. Signal the information that we get from the platoon leader, and then we have some process noise that wk in there. It's just a reduction and available. That accounts for the fact that when you're implementing some of these modeling approach, you might have small adverse effect if you're dealing with the physics of the environment, right? So you just have these small noise that we need to account for.
30:21:705Alessandro Brighente: And
30:24:159Alessandro Brighente: here you have the distributions of all of these variables. Right? Both. Wk is in the process noise, and we said the the measurement noise, same factor. But in here, basically, we have the prediction of the the measurements that we perform on the vehicle.
30:42:870Alessandro Brighente: And then, what is our objective right? So we said that the the noisy component is on top of the information that we receive from the gets propagated in the control strategies of all the vehicles right to get the same here, all of them will contain 0
31:02:140Alessandro Brighente: good. So our objective is to estimate the acceleration of the I minus one vehicle.
31:10:588Alessandro Brighente: Given as input these noisy signal, right? So we assume that we know that the presence of this noisy signal, the idea is that, given that we know which specific, noisy signal has been generated by a platoon leader at a specific time. We also know how it propagates in the control strategies.
31:35:359Alessandro Brighente: And eventually we will find these this signature, this noisy signal also. In in our predictions of acceleration.
31:53:870Alessandro Brighente: Okay, so here's a a picture on how we can. we can represent these this system. Okay, so what happens is that
32:04:580Alessandro Brighente: the the platform leader in here is designing it. So control right? It has no problem. Leader doesn't need to receive information from one other details. It's just a generator of everything great. So the the platform leader has its own say, right decay, and the K. And sends all of this information back to the to the other details. Right?
32:27:110Alessandro Brighente: Good. So the vehicle one received the information from the platform leader, meaning, Dk, 0, ak, 0, and then the Ak 0 right, which is the the the noisy signal propagated by the leader performs. It's it's computations, and then sends its information back to the other vehicles, right? So the I minus one vehicle.
32:51:20Alessandro Brighente: which kind of information we receive, whether we received the status or the I managed to be good.
32:56:390Alessandro Brighente: together with this delta a 0 which, again, is the signature that the platoon leader generated and then propagates its own information back to the other vehicles up to the to the last one
33:06:560Alessandro Brighente: great. So what happens at the I vehicle? Right? So we have on board sensors which gives us information on distance, the speed and acceleration like the thing that the vehicle can actually measure. And from the network we get the information from other vehicles, right? So I minus one for speed and acceleration, and the Delta ap 0, which is the signature embedded originally by the meter
33:31:883Alessandro Brighente: in the control signal right? So the thing here is, we have our controller right? The controller performs. It's a
33:41:630Alessandro Brighente: updates, right? It sends whatever control strategy to the actuator into the network. And then in this finding here, it says we have the detector. Right? So what does the detector do? Whether the detector is trying to understand whether
33:57:290Alessandro Brighente: the values that we received from a nectar actually contain a visa signature in here. Right? So you see one very fundamental point. So this decay and Ak are a result of the fact that these 10 K has been embedded in the same equation right? And so, because
34:18:110Alessandro Brighente: we received this information from the from the leader, and we'll check whether it's prediction of the leader's behavior contain exactly this thing here.
34:26:230Alessandro Brighente: No, the point is, if we receive an update right. A message that has been replayed. The point is that it will not contain a valid signature, right? So there will be a mismatch between our predictions, based on the knowledge, on the signature and the the values that we receive from from the from the communication right
34:51:50Alessandro Brighente: from the docker.
34:53:420Alessandro Brighente: Okay, again. It's not the fact that we are receiving this Delta K, but it's the fact that the Delta K has an impact on the control strategies of all the vehicles. Right? So if we have a replay attack.
35:11:360Alessandro Brighente: then would there would certainly be a mismatch between our predictions and the information that we get from from the people, from from the attacker itself.
35:26:930Alessandro Brighente: Good. So how do we detect for the presence of these? these authentication signal in the control strategy. Right? We we receive these VA, and then the values right? And we want to check whether this data is presence in is present in the acceleration strategy, right in the acceleration signal
35:50:957Alessandro Brighente: in order to check for the presence of signals. Usually we? We run cross correlators. Right? So you have these functions that check for the presence of a specific sequence in a in a time series.
36:04:880Alessandro Brighente: Okay, so how does cross correlation look like? So this is the equation for cross correlation. Right? So you have 2 functions F and G, which are lambda t in here. Right? And you just compute this integral in here. Of course, you will not compute the integral on computers. You will resort to this discrete time
36:27:523Alessandro Brighente: version of this equation. But basically we tell you, whether the signature is present or not.
36:35:640Alessandro Brighente: right? So if the signature is not present, the cross correlation would be 0 if instead, there's a
36:42:190Alessandro Brighente: the signature is there, and you will see higher correlation. Right? Yes, I have an example in here. So
36:51:750Alessandro Brighente: let me just show you.
36:54:650Alessandro Brighente: Yeah, I have it here. So
36:56:890Alessandro Brighente: you see. Yes. So let's say that we have. We have our control singer right? Which
37:05:820Alessandro Brighente: has these time steps in here from 0 to one with that time step.
37:11:670Alessandro Brighente: And then let's say that the transmitted signal is just a sine wave. Okay? And then what do we do in this code? I will put the code online. And it's just very simple, just to give you an idea. So the thing here is we generate the noisy signal, and it's just random noise. Gaussian distributed white noise
37:34:647Alessandro Brighente: and we generate 2 versions of that right? So the idea is that the 1st row is the the legitimate signal, and the second row is the replace signal, right? So we want to detect for the presence of the the 1st signal. Right? So we have 2 signals in here, the legitimate one which is the transmitted signal. So the sine wave, plus the noise at this time.
37:56:690Alessandro Brighente: and the repeat signal, which again might be the translated signal plus a previous version of the noise. Right? And here we're just computing the cross correlation between the the signals and the legitimate noise sequence. Right? So the legitimate noise. Sequence, we said, is the 1st row of these this matrix in here.
38:18:810Alessandro Brighente: and we're just not computing the cross correlation between the replace signal and the noise and the replace signals and the noise, because we want to see where. What is the effect of having a specific signature in our signal.
38:33:30Alessandro Brighente: Okay, so what you have in here is the result
38:38:245Alessandro Brighente: the difference from the figure on the left, on the right and the right is just the fact that this is a singular realization. Right? So we have a 1 realization of the noisy signal, both for the replay one.
38:51:640Alessandro Brighente: And what you see is that. If we compute the cross correlation between the the valid signal and the the actual signature, you get this spike in here, and it means that you actually detected something right? If instead, you you compute the cross correlation between the replay, the signal, and the the legit version of the noise, you see something that is definitely lower
39:20:560Alessandro Brighente: and live.
39:22:20Alessandro Brighente: Okay. So again, this is a single realization. But if you perform this thing over many, many realizations of the noise this is what happens right? The the cross correlation between the signal with the legitimate signature and the signature will have a high peak in here instead, the cross correlation between the replace signal and the legitimate signature will always be close to 0, right? It means that there's no legitimate signal in there.
39:51:720Alessandro Brighente: So this helps you detecting whether the the input that you received, it's actually matching what you expect from the from the valid signature. And so here you're just playing with the with signals one of the things is
40:07:700Alessandro Brighente: now, you see, all the control strategies that we defined up to now seem to be a single value. Right? It's something like you have your linear system. It gives you a prediction of a value, and that would be so. If that would be the case, it would make no sense to have this kind of signals.
40:25:372Alessandro Brighente: What happens in reality is that. We always work with digital systems. But then in the end, we need to provide an analog signal to the actuators that we use. Right? For instance, electronic speed control or something like that. It needs to have a
40:41:690Alessandro Brighente: a continuous time signal, right? So our control strategy will be converted into something that is an analog signal which is definitely not just a single value. But it's over a time series, right? And that's why it makes sense to embed these these kind of signatures
40:58:20Alessandro Brighente: in there.
41:04:160Alessandro Brighente: let's have a break.
41:28:850Alessandro Brighente: Okay.
41:32:930Alessandro Brighente: well, all these things that the autonomous people do are based on on the improvement of things. Right? So up to. Now we we said that the vehicles can sense something right? They can use some distance sensors they can use. Communication means. So they need to understand what the what is happening, right? They need to understand whether they're getting too close to the people in front, for instance.
41:59:300Alessandro Brighente: and they need to act based on that right. So the the control strategy needs to adjust the
42:04:750Alessandro Brighente: the, the distance from the people in the phone itself.
42:08:120Alessandro Brighente: Good. So this has a fundamental assumption of the fact that we are actually able to sense something right? So as we mentioned before, the sensing layer is composed of different kinds of sensors. And so we have lighters. We have radars. We have cameras, right? All these devices that gives information on what is happening. What is in the surrounding
42:31:953Alessandro Brighente: of the vehicle. So in acc, yeah, in acc and cacc, for instance, we said that we have this distance sensor, and they they tell us, whether there's an object at a given distance from our vehicle, and we can estimate the distance.
42:54:120Alessandro Brighente: great. So
42:56:20Alessandro Brighente: these sensors can have, of course, different
43:01:410Alessandro Brighente: reasons for being there. Either. Different functions
43:05:143Alessandro Brighente: they can be used for safety. They can be used for diagnostic right? So the the tire pressure monitoring system. That's a sensor, right? It's sensing the pressures of the tire and telling us, what is the value and see whether it's safe or not to drive? They can monitor the environment, and we can see whether there's a a pedestrian on the road.
43:27:956Alessandro Brighente: But now let's so you have some examples in here on. Why these sensors are important what they might use for. But our problem now is, are there any specific attacks that we can answer against sensor right? Can we impair the sensing capabilities of the vehicle.
43:46:880Alessandro Brighente: If that is the case. How can we do that? Alright? How can we achieve the the point where the sensor doesn't work anymore? Or we let the car or the vehicle, whatever it is. To sensor something that we control. How can we measure, for instance, the the distance measure by the Ccc sensor.
44:09:500Alessandro Brighente: Good. So
44:12:770Alessandro Brighente: When we talk about the sensing layer we have different attack vectors. Right? So one of them is tampering the sensors right physically, changing some of the settings of of the sensor. Right
44:27:236Alessandro Brighente: then stupid. So this would be. I disconnect the wire right? If I disconnect the wire sensor is not able to do its job anymore. But then, of course, that's
44:38:770Alessandro Brighente: not as convenient as an attack. I mean, it's effective. But there's better thing that we can do or not. Necessarily. We have the the capabilities to go there and tamper physically. Tamper the sensors right? So we would like to do to perform some of these attacks from from a given distance, right with certain distance.
45:01:849Alessandro Brighente: So, for instance, we can be on the roadside, and we can have laser pointers or something that impairs the capability of the sensor. We can be another vehicle in the, in, the, in the network or the transportation system, right? And send these issues. But yeah, anyhow, the the point here will, will be related to how
45:24:810Alessandro Brighente: close can we get to the target people? What the reference system I'm using with respect to the the victim vehicle. Right now we can position myself in a place that is effective for the attack, or if I cannot be there, in the most effective locations for the for the attacker.
45:42:210Alessandro Brighente: What kind of effects can can I achieve?
45:48:440Alessandro Brighente: And so we start with the with lighter.
45:51:750Alessandro Brighente: because it's 1 of the most used and studies studied sensing
45:58:320Alessandro Brighente: means for people. Right? So lighter stands for light detection and ranging. And basically, it's a technology that that uses electromagnetic waves to detect for the presence of objects in the street, right in the surrounding from the car.
46:14:580Alessandro Brighente: Good. So we have a active sensing in here we have an active sensor, right? The lighter itself is sending some signals and waits for for the response, and based on the response it can perform some some computation.
46:28:490Alessandro Brighente: Good. So here you have an example of what you detect with the with the line, right? So that's a possible visualization of
46:38:400Alessandro Brighente: of the the points that we get with the lighter. So basically, you have an object on top of the car and then sending these signals measuring the the responses and measuring basically the distance. Right? You measure how much time the signal takes to to get from the lighter to the reflected object and back the lighter. So basically construct basically a map
47:02:320Alessandro Brighente: of the of the road. In this case, right? So you can detect the presence of trees, you can detect the presence of other vehicles on the road. You can detect the presence of walls or the sidewalks right? Everything that basically has a 3D shape and can reflect signals.
47:19:230Alessandro Brighente: Good. So how do lighters work? We have, 2 main types. And the 1st one is the scanning type. And you have this device. Which?
47:34:720Alessandro Brighente: so you have a basis. And then on top of this basis, you have a device that is rotating right on the on the vertical axis, and is sending these signals in all the different directions. Right? 360 degrees.
47:47:90Alessandro Brighente: And it's it's there for moving, right? So for each direction, or with the specific periodicity, we will send a signal. And thanks to this, you are able to basically cover the whole surrounding of the car. So, for instance, in this image in here you have a device that is rotating on top of the car. And so it's sending signals in each of these directions and the recording measurements
48:11:290Alessandro Brighente: I know recording responses with the guests
48:14:394Alessandro Brighente: the second type is a solid state lighters which do not require these moving parts. So it's something like an omnidirectional antenna right? At the same time it can send signals in different directions without moving anything.
48:29:370Alessandro Brighente: The problem with soil state lighters is that they're more expensive. So usually we do not have them. We have. We still have a new stunning lighters right? And we focus on that. We will see how we can deliver attacks against these stunning lighters.
48:47:520Alessandro Brighente: So we said that these lighters are sending. Are actively sending sameness. Right? And yeah, this is very, basically how the idea of the lighter, how that works. Right? So you have the transmit device in here, which is the lighter it transmits a laser pulse right? This laser pulse gets to the object at the time you want
49:10:450Alessandro Brighente: the the laser pulse is reflected by the object. It gets back and gets back to the lighter right, and we know that these timing here the course between the transmission of the laser pulse and the exception of the reflected pulse, is this here? And yeah, again, some basics in here. So you know that the the laser pulse is traveling at the speed of light.
49:35:280Alessandro Brighente: and you can basically measure the distance that the signal traveled back and forth right? So here you have divided by 2, because, of course, you have 2 times the distance. If you're covered the whole time.
49:49:200Alessandro Brighente: transmit you, and then get it back.
49:53:336Alessandro Brighente: Good. So
49:55:520Alessandro Brighente: what the riders do.
49:58:530Alessandro Brighente: and just we ignore the current move because it's so slowly compared to the
50:04:140Alessandro Brighente: yes, yes, here, yes, let's stick with the with the simple thing. So what will happen is that due to movement? You have different effects on the transmitive signals, right but the assumption is that that these effects would be
50:21:690Alessandro Brighente: kind of small compared to the way the signal behaves such that these holds with the given approximation, which is sufficient.
50:34:761Alessandro Brighente: Yeah, good. So we have these devices, this rotating in different direction and sending the signal performing this measurement for for all given set of directions. Of course, if you're pointing your signal at the given height, right then you just over a single plane.
50:54:850Alessandro Brighente: So not only you want to send the signal in all the different directions, but you also want to control what is called the slant angle, right? So this direction here the vertical angle. So not only you rotate in this direction, but you also go up and down to to cover the whole points right? So otherwise you could not get this thing here right. If you just select the again angle, you just have a a single or one of these lines. You said, you want to cover them all
51:26:412Alessandro Brighente: good. So what happens now? You're are sending multiple pulses right for each of these directions, so there might be ambiguities right you're sending a signal that you want to wait for the signal to be reflected and get back to your lighter. But you might have also other signals like, you're sending multiple signals, for instance, how can you be sure that
51:52:100Alessandro Brighente: the one that you're receiving is actually based on the last measurement you want to perform.
51:58:676Alessandro Brighente: So lighter has certain parameters.
52:04:00Alessandro Brighente: Right? So one of these parameters is the receiving time right in. It defines basically, how much time the lighter can wait for the reception of the response signal.
52:17:10Alessandro Brighente: And now we have a better time. Right? So that time means how much time we wait. Between transmitting what pause and the successive one.
52:31:500Alessandro Brighente: okay, so we have a figure for this. Yes, so this is what happens. Right? So you have. Your policy. This is the 1st policy. Right? You you send your signal, and then you wait for the response.
52:43:690Alessandro Brighente: You wait for the response for these that the Max value in here right? So you give time, the signal to travel back and forth for these maximum of time. So whatever you receive in here, you leave it as well, and then you have the time. Right? So how much time do I want to wait before transmitting the next pause?
53:04:580Alessandro Brighente: Right? So everything that I received in the receiving time? I think it is valid. If I receive something here in that time. I do not give it as valid right, because it doesn't make sense with the with the parameters of my license right? This is something that we use to to avoid getting noisy values, or something that is not accurate. Right? We say that if we receive something in this that time in
53:29:510Alessandro Brighente: it might not be related with the latest pulse that we translate to. Okay?
53:35:860Alessandro Brighente: Good. So it means that. If we set the a maximum receiving time, it means that we can measure distances
53:44:74Alessandro Brighente: within a specific range, right? And this specific range is given by again this competition here. So with this Slider, right? Setting a specific density. Max, we can measure objects that are at maximum the distance and mask here.
54:02:990Alessandro Brighente: Okay, so another parameter that we have is these false repetition time? Right? So how much time does it incur between a pulse and the other right? So this defines the periodicity of the policies that the the lighter sends. Right? So you are sending signals at these specific areas. You're not that you're going to just get a signal every
54:27:470Alessandro Brighente: the capital 3 segments in here.
54:34:37Alessandro Brighente: Not only
54:38:270Alessandro Brighente: So let's put it this way. If you are transmitting a signal right? And you want to to receive the signal back.
54:46:10Alessandro Brighente: How do these things work? Right? It's not like you have just a specific direction, and you're able to receive signals only from that specific direction. Or, if that's the case.
54:55:360Alessandro Brighente: you're limiting to match the capabilities of your lighters, and then we need to take into account the fact that the lighter is rotating while doing this kind of stuff right? So if it's sending a signal towards the direction and then rotates a bit. It needs to be able to receive the signal back from from that direction. So we have something that is called the aperture of the lighter. It basically tells you the
55:18:870Alessandro Brighente: the the width of the angle in which the lighter can receive signals back right? So you have this equation here that tells you the the receiving angle. Right? So this receiving angle is given by
55:35:19Alessandro Brighente: these, then the p maximum here times the the speed at which the lighter rotates. Okay? And you basically can compute how much?
55:45:920Alessandro Brighente: which angle basically, you cover at a given time. Right?
55:51:10Alessandro Brighente: Delta. Max, again, is the time in which we dim, I suppose, as well. Right? So if we transmit a certain signal in that direction again. I'm rotating at a given speed. I know that I can still receive a signal they sent from that direction after Delta Max. So where's my point in direction after Delta Max.
56:15:220Alessandro Brighente: And this is exactly what you're having here. So you need to take into account for that angle and have something that's able to receive signals up to that angle.
56:27:170Alessandro Brighente: And of course you can control this with the with the rotation speed and the with the the best remarks that you want to set right. So the receiving time.
56:36:990Alessandro Brighente: Good! How does the sensing capability of the lighter works? Right? This is something that holds through generally for for different type of sensors.
56:45:746Alessandro Brighente: So you have. this kind of behavior for sensors, right? So you receive an input, right? Your input in this case is the laser pause that gets back from the from the reflected objects. Right? And and these inputs triggers a reaction from the from the circuit point of view, and it triggers a current value, for instance, a current signal.
57:08:680Alessandro Brighente: And so in this figure you have how the output of the sensor behaves. Given different inputs. Right? So the input in here is increasing. It's something like you're receiving a laser puls with an increasing power. Right? And this is what happens. So if you're receiving power
57:30:930Alessandro Brighente: is too small, then you have this sided region here. It doesn't trigger any output the sense, right? So the sensor is not sensing
57:41:19Alessandro Brighente: anything. It's something like when you're in a room that is too dark, right, and if there's no light source, you're not able to see anything. That's exactly how this thing working here. Right? So here there's no light strong enough to to trigger a reaction in the sense.
57:59:180Alessandro Brighente: Then you have this thing here, which is the linear region? Right? So if the the power of the same signal keeps increasing. Then you trigger an output of the sensor. Right? It's something like you see that the example of the room, the dark room that we had before. If you have a light source with increasing power, then you start seeing more and more, and realize that there's objects and define
58:21:828Alessandro Brighente: objectively, we see their shapes and colors and stuff. I think this is exactly what happens with their vision.
58:27:560Alessandro Brighente: And then, at a certain point, you reach these these
58:32:610Alessandro Brighente: region here which is for saturation. Reason, right? It's something like if the light source in your dark room is too strong, then you're not able to see anything anymore. Right? You just see huge light source it's exactly what happens when we send here. And we'll see that
58:51:930Alessandro Brighente: right? This is the the separation region. So the the sensor is not able to sense anything anymore. Right? So assuming that we have 2 light sources, right if one is strong enough to cause the the sensors to transition to the separation region.
59:07:720Alessandro Brighente: I am not able to see anything else. Right? So if there's a a smaller power source. I don't see it. I don't see anything. Okay. And
59:18:640Alessandro Brighente: good. This gives us, a 1st intuition how we can attack this kind of sensor right. If we have these behaviors for sensors we can simply attack it.
59:31:610Alessandro Brighente: driving them to the circulations agent. So if I point the something at the later
59:36:170Alessandro Brighente: that is too strong. Basically, the lighter is not able to censor anything anymore. Right? So it will be sending its legitimate signals back and forth in all the different direction, but will just sense my attackers stronger signal right? And so basically, we are preventing the lighter from collecting its point.
59:58:230Alessandro Brighente: yeah, it's cloud of points that define. The objects in the in the environment.
00:04:660Alessandro Brighente: Can this happen? Even if you have 2 liners?
00:07:720Alessandro Brighente: I mean, one is sending, and the other one is sending too.
00:12:60Alessandro Brighente: Yes.
00:14:60Alessandro Brighente: it depends on how close they are right, because if you want to bring the the sensus to the separation reason, the signal that it receives needs to be strong enough.
00:22:960Alessandro Brighente: So if the lighters at the same time have a
00:28:46Alessandro Brighente: the the angle that we've seen before, right? They need to be overlapped somehow. And in that case, if one of the lighters is pointing towards an angle that the other one is able to censor, and is close enough or strong enough
00:43:70Alessandro Brighente: then. Yes, because I was thinking about the the example you you showed before the car. If we think about a couple of cars, each one with its own lighter, and you have a lot of them. But they are separated enough not to one another. But
01:02:700Alessandro Brighente: if the attacker is one of the vehicles of the platoon, and he's using a lighter that is more powerful than it's supposed to be then. Yes, that's exactly what happens.
01:16:430Alessandro Brighente: Good. So these this situation attack is basically a denial of service attack. Right? So what's happening in this situation again. We are we already since this, but we are exploiting somehow a semantic gap that exists between the real world and the world says by the lighter. So what does the the lighter detect in this situation? Well, if the lighter detects super strong signal
01:43:143Alessandro Brighente: then it thinks that it's very close to. To a possible reflection source, right? The fact of having something that is very powerful is something like, I'm transmitted a signal. The signal power did not decrease enough in traveling. Right? So it means that it it was kind of very fast.
02:01:770Alessandro Brighente: and that's 1 of the effects, or the other effect is simply the the blinding that we see before right the fact that if I have a very strong source, then I just don't see anything anymore.
02:17:510Alessandro Brighente: Okay, so yeah, here you have another example of what the semantic gap means. Right? If we have a seismometer, and then we want to measure an earthquake. The seismometer doesn't know whether there's an actual earthquake or a child shaping it right in this. Exactly what is happening here with the withers.
02:40:390Alessandro Brighente: So
02:42:170Alessandro Brighente: is it really that simple? Well, lighters do not necessarily
02:49:310Alessandro Brighente: accept all the the waveform that they receive right, because otherwise it will be
02:56:500Alessandro Brighente: somehow weird
02:58:584Alessandro Brighente: somehow weird in the sense that there might be strong light sources, or there might be strong seen as a delighter. We'll proceed, not due to attackers, but due to different circumstances. So what they did is designing lighters such that they they send
03:20:400Alessandro Brighente: signals with a specific shape, right with specific waveforms, and the only accept the responses with that specific waveform. Right? The point is that if I'm sending a waveform, and it gets reflected on an object the waveform that is reflected back will not change its shape right? So the lighter will detect whether the signal it received matches its own expected waveform.
03:47:566Alessandro Brighente: Great. So somehow this is similar to what we've seen now with the keys and the cards right? We are expecting something, but it means that the attack that we had the 4 keys of work works exactly the same in here.
04:02:920Alessandro Brighente: Right? So it means that, for instance, we can either capture the the waveform that the lighter sends right and have that specific waveform in our pocket, and whenever we want to check the lighter, whenever we want blind lighter, we'll just replay that exact waveform with the stronger power, right? So that the the lighter will receive this waveform see that it's a valid one, and accept it right, and pass it through the
04:29:790Alessandro Brighente: the chain breaks into the to the sense that.
04:35:00Alessandro Brighente: And again, here we have something that is difficult to detect. Right? How can we detect whether a signal with a specific waveform has been generated by a reflection on an object, or whether it has been replayed by an attacker that just captured it, and send it back to the to the lighter. So this is kind of
04:54:940Alessandro Brighente: this is actually a powerful effect. It still works. Okay, so we want to run these these blinding attacks. Okay? And, as we said before, we want to point a laser towards and lighter such that we bring it to the service, right? So light source is strong enough to bring lighter sensors to the saturation agent.
05:21:557Alessandro Brighente: Good. So we need to to deal with some characteristics on the data right? So how these sensors are made, or where they are placed, or the fact that they rotate
05:33:110Alessandro Brighente: alright. So what are the different common features that we have for for blinding attack. Well, the 1st thing is the the steadiness of these attacks. Right? So
05:43:908Alessandro Brighente: imagine something. Imagine that you have a lighters right on top of cars, and they're driving around, and they're just sending light beings in each direction. You will actually blind the drivers like that would not be good. So what we use for lighters is these infrared lasers. So it means that you're not able to see the signals that the lighters are sending right.
06:05:50Alessandro Brighente: And so it means that if we have, if we consider an attacker that is launching these blinding attacks, we are not able to see the the attacker scene, and we are not able to.
06:16:620Alessandro Brighente: As humans, we are not able to see infrared
06:21:329Alessandro Brighente: light sources, right? So we do not know whether there's an attacker launching a signal from a manual direction at least as drivers.
06:32:877Alessandro Brighente: Then, receiving angles. Right? We said that we have these devices, that rotating that have a specific receiving angle that has some parameters
06:43:437Alessandro Brighente: and we might think that in order to blind the riders. We we need the victim to have a very wide receiving
06:53:960Alessandro Brighente: and good. But that's actually not the case
06:58:528Alessandro Brighente: usually the receiving angle of the lighters is kind of small
07:05:124Alessandro Brighente: and somehow it limits the effect on the situation. Right? So you have a small receiving angles, and these devices is rotating. So this means that you will be able to blind the lighter just in a given direction. Right? So if you're pointing your strong, light source towards the lighter from a given angle. Then the ladder will not see anything in that specific direction, only it's a receiving angle, which is.
07:31:990Alessandro Brighente: I mean, we're not completely
07:36:160Alessandro Brighente: blocking the ladder from sensing anything
07:39:675Alessandro Brighente: in the world, but just in a specific region. But
07:43:270Alessandro Brighente: good enough
07:44:867Alessandro Brighente: and then the receiving angle. Right? So usually, if you take a look on how a liner is structured again, you have the basis. You have the the rotating object in here descending and receiving angles, and on top of that you have a reception glass. Imagine it has the what you have on cameras. Right? You have this curved glass on top of the line. So it means that the the
08:11:720Alessandro Brighente: the signal that impairs on the curve. The reception glass might be rotated right? So if you have something that is pointing directly towards the glass, and it's fine, it will go straight
08:23:800Alessandro Brighente: into the liner in the structure. But if you said that you're sending your signal from which is not a straight line, this would be heard right. And here you have 2 examples. So it means that from these light source in here you can induce paper measurements of the liner on that direction.
08:44:939Alessandro Brighente: And in here is that when you have multiple reflections inside the curve glass, although your light source in here isn't here, you will take those fake measurements from the ladder from these directions here, and so we need to take that into account as well. Either we directly point towards the liner, or we'll cause blindings in different directions.
09:13:10Alessandro Brighente: Good. So here you have, some numbers. Right? So what do we need? Actually, in order to saturate the lighters. So you have some example in here. 30 millivolt
09:27:410Alessandro Brighente: 905 nanometer laser module, which is $40
09:34:384Alessandro Brighente: or another thing here, and 905 nanometer laser module, $350. Both of them work to saturate the lighters. Right? So you can have these devices, point them against the lighter and get the saturation, the successful blinding attacks.
09:55:780Alessandro Brighente: So what is the effect of these situations? Right? We can consider 2 2 different scenarios the 1st one is the one here on the left, where we have a a week light source.
10:06:470Alessandro Brighente: right? So we are sending our signals towards a lighter, and then the lighter will detect some dots in here. Right? So we're just sending our replayed value here, and we create the these dots right? So there will be kind of random. I mean, we are not controlling anything at this point. We're just sending signals valid signals back
10:30:870Alessandro Brighente: to the lighter. And if this is the direction of the light source, then we're sending something to the lighter. You actually will have a certain receiving angle, and we'll detect this here.
10:44:302Alessandro Brighente: If instead, we consider
10:48:612Alessandro Brighente: a stronger, oblique light source, it means that it's not pointing directly at the the center of the light of it is a bit
11:02:870Alessandro Brighente: unaligned
11:04:657Alessandro Brighente: these this signal will be curved up by the, and these is what you get in here right? So these will be the direction of your light source, but due to the parameters that you have in the glass. It will detect these dots in here. So these are experiments performed on the on the actual laser. Right? So you you send your your signal in that direction, and you can pick dots.
11:27:690Alessandro Brighente: So you see that this is very strong from an attacker's point of view, because it means that you are. You're sending something from a given direction. And then you're posing artifacts in a completely different direction. So if the big thing expects
11:43:315Alessandro Brighente: to find the attack in here. Well, actually, not right. It's just an effect that you have because of the parameters that you have on the glass of the of the lighter.
11:55:290Alessandro Brighente: And what happens instead with the very strong, light source. Right now, we have something that is actually able to saturate the the lighter.
12:04:560Alessandro Brighente: So these are the points of the line of the tax before the attack. Right? So you see that the
12:11:360Alessandro Brighente: you have all this option here, meaning that it gets reflections from these points. And after the fact, you see this thing in here, right? So in here you have something that prevents the ladder from measuring anything from this direction, and you see the 3rd part canceling the artifacts that you had before. So it means, for instance, that if these would be a sidewalk or a traffic sign, or whatever that could prevent the the record from crashing into something
12:41:42Alessandro Brighente: you're now cancelling it right? So the the vehicle is not able to detect it anymore, and might crash into that, or at least not consider the fact that it needs to avoid to go in that direction. Okay, so this.
12:56:980Alessandro Brighente: this is actually the factor of line. Right? You don't see these artifacts anymore. And that's why it's for the lining. Again. In this case you have something that is pointing a very strong like source towards the ladder in this specific direction, right? And the ladder is not able to to get any responses back from
13:18:180Alessandro Brighente: from that direction.
13:22:680Alessandro Brighente: Let's see, okay.