1
00:00:03,430 --> 00:00:05,950
Good day, DEF CON and Car Hacking Village.

2
00:00:05,950 --> 00:00:11,550
I am Jonathan Petit, and I'm here with two of my colleagues, Rashid Ansari and Kong Chen.

3
00:00:11,550 --> 00:00:17,430
We are the research team for Qualcomm working on connected and automated vehicle security.

4
00:00:18,150 --> 00:00:26,730
Today, we're going to talk about misbehavior detection and demonstrating you some attacks on VLAN board unit.

5
00:00:27,010 --> 00:00:32,230
So before we start, I think it is important that we understand what is V2X communication.

6
00:00:32,230 --> 00:00:50,070
So V2X enable devices like vehicles, pedestrians or roadside unit to be equipped with an onboard unit, a device that enables all of those components to broadcast information in the surrounding to create an awareness.

7
00:00:50,330 --> 00:00:57,850
So in this example, we have three cars and are all equipped with V2X technology.

8
00:00:58,190 --> 00:01:02,530
And we'll see that now the white car is performing an emergency brake.

9
00:01:02,890 --> 00:01:16,390
Therefore, because the onboard unit is connected to the CAN bus, it can detect that there is a high deceleration, for example, and that in the brake pressure, and therefore we'll be able to generate a message that he's sending to the surrounding vehicle.

10
00:01:17,150 --> 00:01:23,750
So those messages are called basic safety message here in the US, so that's the BSM that we're going to talk about in this talk.

11
00:01:24,390 --> 00:01:35,070
When the vehicle receive a BSM, they are able to understand that the white car is performing an emergency brake, and therefore themselves brake, avoiding a chain of collision.

12
00:01:35,130 --> 00:01:41,890
So we see the huge benefit, safety benefit, especially in that V2V use case that we are showing you.

13
00:01:41,890 --> 00:01:51,350
This is an emergency electronic brake light application, and it's consuming BSM sent by other vehicle to understand an emergency brake is happening.

14
00:01:51,430 --> 00:01:56,330
So there is many application for safety and auto efficiency that are designed for V2X.

15
00:01:56,330 --> 00:01:57,710
So this is the context here.

16
00:01:57,710 --> 00:02:00,490
We are looking at those type of application.

17
00:02:00,730 --> 00:02:12,030
So then, of course, we understand that from a security standpoint, we do not want external attacker to be able to send bad messages to the other vehicles.

18
00:02:12,030 --> 00:02:16,330
So all the messages that are sent are actually digitally signed.

19
00:02:16,330 --> 00:02:27,090
So in the same example we just showed before, now we have the white car is actually signing with this private key, attaching a certificate, the message before sending it.

20
00:02:27,090 --> 00:02:34,890
The receiver will use the keys in order to verify that this is sent by an authenticated vehicle.

21
00:02:35,210 --> 00:02:36,350
Good.

22
00:02:36,610 --> 00:02:43,370
But as you might imagine, this doesn't prevent any bad content to be sent.

23
00:02:43,370 --> 00:02:47,670
You could be authenticated, but sending wrong information.

24
00:02:48,190 --> 00:02:55,270
So here in our same example, we have here the white car is an attacker, is sending a fake EBL to the other vehicles.

25
00:02:55,610 --> 00:02:58,310
The other vehicle will verify the signature.

26
00:02:58,310 --> 00:02:59,530
It's all valid.

27
00:02:59,570 --> 00:03:03,990
Therefore, they have to use a content which says it is an emergency brake.

28
00:03:04,310 --> 00:03:08,990
Here in the animation you just saw here, the blue car didn't stop.

29
00:03:08,990 --> 00:03:10,010
And you might wonder why.

30
00:03:10,010 --> 00:03:17,890
It's because we can make an assumption that this vehicle was equipped or is equipped with a camera and therefore didn't see the white car braking.

31
00:03:18,190 --> 00:03:22,970
But the red car has no line of sight and therefore has to trust that information.

32
00:03:23,550 --> 00:03:30,310
So we clearly understand and you need a misbehavior protection system to defend against such attacks.

33
00:03:30,370 --> 00:03:39,670
And this is one we're going to present today is really about the attacks and how you perform this and how you protect against this.

34
00:03:40,070 --> 00:03:45,770
So let me show you a little bit on our example about misbehavior.

35
00:03:46,150 --> 00:03:52,570
Here on the left-hand side, this is an application called intelligent signaling.

36
00:03:52,830 --> 00:04:00,730
This is where you have a traffic light equipped with a roadside unit with the same kind of equipment to receive the message sent by the vehicle.

37
00:04:01,570 --> 00:04:07,590
And by reading the BSM, it adjusts the signal phase and time.

38
00:04:07,590 --> 00:04:12,710
Therefore, it's going to play with the red and green on this different segment coming at the intersection.

39
00:04:13,490 --> 00:04:32,530
And an attack that was demonstrated by some researcher showed that if the vehicles are sending wrong BSMs, I'm going to just run the video here, if you send that, and after 30 minutes of attack, you can see this long line at each of two, or actually three intersections,

40
00:04:32,530 --> 00:04:34,070
three segment of the intersection.

41
00:04:34,070 --> 00:04:46,690
So an application that consumed a BSM that is fake, even though it's purely authenticated, if you do not do a misbehavior analysis protection, you might see issues like this in intelligent signaling.

42
00:04:47,890 --> 00:04:53,070
Now looking on the right-hand side, we have a position jump attack.

43
00:04:53,290 --> 00:05:00,110
So the position here in that representation, you are, the ego vehicle is the green dot.

44
00:05:00,110 --> 00:05:09,690
You're driving and an attacker is sending messages with the same identifier, but he's jumping, he's changing his position.

45
00:05:09,690 --> 00:05:13,090
That's where you see this blue dot, this is the BSM received.

46
00:05:13,090 --> 00:05:16,670
And it's flickering because it's actually jumping back and forth.

47
00:05:16,770 --> 00:05:21,830
And sometimes outside of the frame here that you can see, but it's basically doing it back and forth.

48
00:05:21,830 --> 00:05:23,610
That's the position jump attack.

49
00:05:24,350 --> 00:05:34,750
And what is interesting, as you're going to hear this sound in a few seconds, is that sometimes this is performing here a forward collision warning.

50
00:05:34,750 --> 00:05:41,090
So even though this was a position jump, you might see you need to detect those kinds of behavior.

51
00:05:41,090 --> 00:05:43,390
Same identifier and still jumping back and forth.

52
00:05:43,390 --> 00:05:46,750
That shouldn't be taken in consideration by an application.

53
00:05:48,090 --> 00:05:50,510
I have other example to show you.

54
00:05:50,550 --> 00:05:55,150
Here, this is another representation of the onboard unit.

55
00:05:55,150 --> 00:05:58,150
Here, this is, we use MATLAB actually as a visualization.

56
00:05:58,370 --> 00:06:01,030
And you are the red car.

57
00:06:01,110 --> 00:06:09,830
And all the other blocks that you see, the blue and orange one, are the vehicles also sent by the V2X message.

58
00:06:09,830 --> 00:06:13,390
So you're going to represent this local dynamic map with all the vehicles.

59
00:06:13,390 --> 00:06:15,630
On the left-hand side, you have the first misbehavior.

60
00:06:15,630 --> 00:06:17,110
It's changing headings.

61
00:06:17,110 --> 00:06:28,490
Because part of the BESM, there is the position, there is your kinematic states, your velocity, but also you have the headings that helps pass planning, for example.

62
00:06:28,490 --> 00:06:30,650
To know where is this car going?

63
00:06:30,650 --> 00:06:33,630
Especially when you perform pass prediction when you're using Kalman filter.

64
00:06:33,630 --> 00:06:35,190
You want to use that heading.

65
00:06:35,490 --> 00:06:39,810
And here, in this specific example, you see the car actually spinning.

66
00:06:39,810 --> 00:06:41,950
So we are changing the heading all the time.

67
00:06:42,150 --> 00:06:49,210
So now your system that consumed the BESM to perform pass planning might be really confused by this kind of behavior.

68
00:06:50,190 --> 00:06:52,770
In the middle, we have a single attack.

69
00:06:52,870 --> 00:06:58,810
So this is where an attacker is using multiple certificates that are all valid at once.

70
00:06:58,810 --> 00:07:01,610
So it could be 20 certificates valid per week.

71
00:07:01,610 --> 00:07:10,190
And for example, you'll be using your 20 or 19, depends if you want to use yours, and you send your 19 to create 19 other ghost vehicles on the road.

72
00:07:10,190 --> 00:07:12,790
And you can make them move as you wish.

73
00:07:12,790 --> 00:07:21,850
So that is what happens when you map all the BESM you receive and put them on this virtual view.

74
00:07:22,130 --> 00:07:30,710
It is really hard for the connected and even automated vehicle in the future to behave in this kind of really noisy environment.

75
00:07:32,490 --> 00:07:34,790
On the right hand side, we have a position overlap.

76
00:07:34,790 --> 00:07:38,330
This is where an attacker is actually overlapping two of those vehicles.

77
00:07:38,330 --> 00:07:40,530
So here, that is basically what we see.

78
00:07:40,530 --> 00:07:43,370
And here even overlapping with your own position.

79
00:07:43,970 --> 00:07:51,310
And those are just a couple of example about what we call misbehavior that are all authenticated messages, but the content is wrong.

80
00:07:54,330 --> 00:08:00,050
So when we look at V2X misbehavior, there's what we call a misbehavior life cycle.

81
00:08:00,170 --> 00:08:11,310
So you receive those V2X messages, and here we gave example about basic safety message because they are the one commonly frequently sent every 100 millisecond to other neighboring vehicle.

82
00:08:11,310 --> 00:08:13,730
The first stage is the detection.

83
00:08:13,730 --> 00:08:15,550
You want to do misbehavior detection.

84
00:08:15,590 --> 00:08:17,710
So what do you do in this case?

85
00:08:18,270 --> 00:08:26,550
The goals are to validate the position of the sender, his motion, its motion, and the presence.

86
00:08:27,870 --> 00:08:33,190
To do that, you can also, you have to look at the application content.

87
00:08:33,190 --> 00:08:37,610
So the application layer content, meaning you look at the content of the message itself.

88
00:08:37,750 --> 00:08:42,770
You will look at a different field, are they plausible, are they actually consistent?

89
00:08:43,870 --> 00:08:52,370
But one important thing is that you need to understand the effect on the V2X application that is consuming that ESM.

90
00:08:52,570 --> 00:08:57,030
Because not all messages are actually triggering an application event.

91
00:08:57,030 --> 00:09:16,310
In our case of the EEBL that we showed earlier, you will perform an EEBL, so perform yourself a heart breaking or warning the user only if the vehicle is ahead of you with a sudden deceleration, for example, and a brake status flag at one.

92
00:09:16,950 --> 00:09:25,510
So I could be sending wrong messages, but I'm not ahead of you, and therefore it will not trigger the V2X application.

93
00:09:25,510 --> 00:09:30,970
So it is important to understand that because not all messages are equal and therefore should be investigated.

94
00:09:31,410 --> 00:09:39,690
As we're going to demonstrate in the demo later, it is really also important to look at the lower layer of the IETF stack.

95
00:09:39,690 --> 00:09:51,150
I'm going to give you an example where you can use signal strengths of the signal to really identify the position or the presence of another vehicle.

96
00:09:52,050 --> 00:09:57,930
So, but then really finally, when you build a detection system, you need to have a timely detection.

97
00:09:58,170 --> 00:10:02,350
You want a real time detection because this is a safety of life application.

98
00:10:02,350 --> 00:10:07,670
This is the information is useful to, in our example of an EEBL to perform an emergency brake.

99
00:10:07,670 --> 00:10:12,110
Therefore you need to detect that attack really quickly.

100
00:10:12,110 --> 00:10:16,270
You cannot afford to wait the whole chain until the global reaction.

101
00:10:17,330 --> 00:10:38,970
And you have to do this, so in real time, if possible, and with high accuracy, and that's the performance part where in some cases, our case, for example, we want to favor recall over precision because it is important to capture as much attack as possible and really precise only on one attack.

102
00:10:39,250 --> 00:10:43,690
So, but that will depend on your system and your use case and what you will consider.

103
00:10:43,690 --> 00:10:48,370
So this is everything you need to consider when you want to build a detection system.

104
00:10:49,130 --> 00:10:54,350
As soon as you've detected, it's not just enough to detect and give it to the application.

105
00:10:54,350 --> 00:10:56,430
You need to perform a local reaction.

106
00:10:57,290 --> 00:11:05,050
And two example about what could happen locally on your vehicle is to dismiss the further messages from the same identifier.

107
00:11:06,190 --> 00:11:17,050
But as we know, there is privacy by design in V2X, which means that we, each vehicle are using pseudonyms and can change them according to pseudonym change strategies.

108
00:11:17,050 --> 00:11:19,490
So what about pseudonym change in this case?

109
00:11:19,490 --> 00:11:23,850
If you dismiss that, the same standard will simply change identifier.

110
00:11:24,010 --> 00:11:25,630
It's still a question here.

111
00:11:25,690 --> 00:11:30,130
Another reaction that we do locally is to adjust your kinematic state.

112
00:11:30,130 --> 00:11:33,170
So you will perform potentially a graceful degradation.

113
00:11:33,450 --> 00:11:42,410
Let's say that you are in a platoon of vehicle, of connected vehicle, where you use those BSTEM to know that the vehicle is ahead and agree on the inter-vehicle distance.

114
00:11:43,190 --> 00:11:45,990
And now you detect that in that platoon, there is a misbehavior.

115
00:11:46,270 --> 00:12:00,750
You better actually say, oh, let me leave the platoon or degrade and say, if this was a level four autonomous vehicle, for example, you want to go a level three, you want to go back to the driver, or you want to warn the driver.

116
00:12:00,750 --> 00:12:08,130
So it is important to think in terms of the cyberspace, dismissing the further message and also on the physical space, what do you do?

117
00:12:09,890 --> 00:12:15,050
Then also, when you do your detection, you will perform, you will generate a report.

118
00:12:15,050 --> 00:12:16,490
That's the reporting phase.

119
00:12:16,490 --> 00:12:22,670
So here, depending on the type of misbehavior, you will have different type of misbehavior report.

120
00:12:22,930 --> 00:12:28,150
And this is currently under standardization to actually agree about what should be in that report.

121
00:12:28,590 --> 00:12:30,150
When do you put the evidence?

122
00:12:30,150 --> 00:12:31,750
What evidence do you need?

123
00:12:31,750 --> 00:12:39,650
And this report is important because it is sent to encrypted and sent to the MA, the Misbehavior Authority.

124
00:12:39,650 --> 00:12:44,890
It is a component part of the PKI for, and in our case, the SCMS.

125
00:12:45,440 --> 00:12:52,590
The Misbehavior Authority, receive all the misbehavior report from different vehicles, aggregate those misbehavior report.

126
00:12:52,670 --> 00:13:06,330
And it is the only one that can actually link this different pseudonym to the same device by actually a process of talking to the other SCMS component, PKI component.

127
00:13:06,630 --> 00:13:14,550
And this is how, for example, they could detect a Sybil attack because they will know that all those devices are linked to the same pseudonym.

128
00:13:15,090 --> 00:13:18,710
So that is an example about what is happening in the investigation.

129
00:13:18,950 --> 00:13:23,910
And then the MA will decide what is the global reaction.

130
00:13:24,110 --> 00:13:28,290
In cyberspace, this is a revocation, for example, of the enrollment certificate.

131
00:13:28,410 --> 00:13:37,310
If you cannot revoke, as it is the case currently in Europe, you just block the generation of future pseudonym certificate for that vehicle.

132
00:13:37,310 --> 00:13:46,810
And then you will put that on, for example, in the US, we put that on the CRL, the Certification of Location List, that is then provided to all vehicles so that they can dismiss immediately the message.

133
00:13:47,870 --> 00:13:51,710
But something that isn't discussed yet is really what happened in the physical space.

134
00:13:52,010 --> 00:14:03,070
You could, for example, inform the owner of the vehicle, of the device, and to perform an inspection, to repair, and therefore to perform a re-enrollment into the system in a secure location.

135
00:14:03,530 --> 00:14:07,370
Or maybe why not performing a remote vehicle immobilization?

136
00:14:07,370 --> 00:14:11,950
For now, you are stopping the vehicle where it is because that's been a misbehavior.

137
00:14:11,950 --> 00:14:19,170
Of course, this remote vehicle immobilization could be problematic if this is coming from a faulty sensor and not a real malicious attacks.

138
00:14:19,770 --> 00:14:25,190
But you know what we are saying, it's like there's a need of cyber reaction and physical space also reaction.

139
00:14:25,370 --> 00:14:27,630
So this is the whole misbehavior lifecycle.

140
00:14:27,990 --> 00:14:40,270
And as we can definitely see it, as an attacker, if I fool the detection part, you will never report me, you will never react to it, you will know no investigation and no verbal reaction.

141
00:14:40,270 --> 00:14:48,290
Therefore for this presentation, we're going to focus on this part about how an attacker can defeat the detection system.

142
00:14:48,290 --> 00:14:49,610
And we're going to show you a progression.

143
00:14:49,610 --> 00:14:52,230
How do we usually do the attacks?

144
00:14:52,230 --> 00:14:56,170
So where are we going to start and then slowly make it a smarter and smarter attacker.

145
00:14:56,170 --> 00:14:58,790
For this now, I will have my colleague, Rashid.

146
00:14:58,790 --> 00:15:02,790
I'm going to take over to talk to you about this progression of attacks.

147
00:15:17,480 --> 00:15:18,720
Can you hear me now?

148
00:15:25,580 --> 00:15:26,980
All right, yeah.

149
00:15:26,980 --> 00:15:28,120
Thank you, Jonathan.

150
00:15:28,120 --> 00:15:39,080
So as Jonathan talked about the detection portion of misbehavior detection system.

151
00:15:39,220 --> 00:15:49,340
So this slide shows you all the possible ways that we look at right now to detect any attack.

152
00:15:49,340 --> 00:15:59,620
So any attack on the application or on the physical layer, or anything.

153
00:15:59,620 --> 00:16:12,500
So let's go over what we do about, how we go about checking all the, checking everything in the V2X space here.

154
00:16:12,500 --> 00:16:23,560
So first we look at basic plausibility checks, where we look at if the speed or angle or acceleration in a BSM or a CAM message is unrealistic.

155
00:16:23,560 --> 00:16:28,360
Is it too high or is the turning angle too much as the vehicle?

156
00:16:28,360 --> 00:16:33,920
As you saw in the heading change example, is the vehicle really changing its heading a lot?

157
00:16:33,920 --> 00:16:35,200
Is this really possible?

158
00:16:35,200 --> 00:16:38,200
And we can do this on a per message basis.

159
00:16:39,620 --> 00:16:45,020
And that's how, and you don't need a chain of messages for this.

160
00:16:45,840 --> 00:16:51,100
So an attack can be detected with one message as well in this case.

161
00:16:51,380 --> 00:16:53,880
Then we look at consistency with the sensor.

162
00:16:53,880 --> 00:17:14,780
So we look at whatever messages we are getting, like whatever speed you're getting, you compare that with, let's say a radar, like if a radar is tracking any other vehicle, whatever speed that radar, that you can calculate using the radar input, you compare that with the information you're getting in the V2X,

163
00:17:14,780 --> 00:17:16,100
in the BSM.

164
00:17:17,440 --> 00:17:25,300
So that way this can also be done in a per message basis, but can only be done on vehicles that are nearby.

165
00:17:25,300 --> 00:17:31,540
And by nearby, I mean in line of sight, because radars, cameras can only work in line of sight, but not in non-line of sight.

166
00:17:32,120 --> 00:17:42,320
Similarly, we look at the RF, we compare the position, the direction, the velocity, we look at the consistency on the RF side as well as on the V2X side.

167
00:17:42,920 --> 00:17:46,480
And then we would also look at map data.

168
00:17:46,480 --> 00:17:49,680
So we look at the consistency with the map data we had.

169
00:17:50,220 --> 00:18:00,500
If a car, for example, has a position that goes through buildings, that is not really a lane that we know about.

170
00:18:00,540 --> 00:18:13,580
So we would really like to see, okay, is the information coming in the V2X message, does that correlate with what we have in the map data?

171
00:18:13,660 --> 00:18:18,940
And then we look at consistency between the senders' messages.

172
00:18:18,940 --> 00:18:36,160
So consistency within the message itself from a sender, like if the brake status, for example, if the brake status is not consistent with deceleration between the messages, like if the brake is applied, is the acceleration negative?

173
00:18:36,160 --> 00:18:37,920
Is the vehicle really decelerating?

174
00:18:37,960 --> 00:18:43,640
So that is one example we look at, is the position, speed, and acceleration, is that consistent with each other?

175
00:18:43,640 --> 00:18:44,960
So that is one more thing.

176
00:18:44,960 --> 00:18:50,540
And there are many more things that we can look at in a VSM or a CAM message.

177
00:18:51,580 --> 00:18:55,920
Then we look at consistency with other vehicles' messages.

178
00:18:55,920 --> 00:19:14,000
So for example, if you have multiple vehicles going on a highway, and you look at one-way traffic, all the vehicles are heading towards this one direction, and one vehicle is coming directly in the opposite direction.

179
00:19:14,000 --> 00:19:16,620
So that is a possible misbehavior.

180
00:19:17,140 --> 00:19:25,240
It can be that there is some crazy driver who is coming on this direction, but it is really highly unlikely, and we consider that as a misbehavior.

181
00:19:25,240 --> 00:19:33,600
Or we look at, that is one example, but in general terms, we look at the trajectory that is implied by one car and match it with others.

182
00:19:33,740 --> 00:19:41,540
Like for speed, for example, the speed of vehicles should almost be similar on a highway scenario as well.

183
00:19:41,540 --> 00:19:43,120
So we look at that as well.

184
00:19:43,870 --> 00:19:46,800
Now, this is how we detect.

185
00:19:47,060 --> 00:19:51,240
And now talking about the fun part, how do you attack all of this?

186
00:19:51,700 --> 00:20:09,280
So what we do is we look at how an attacker will go about attacking the misbehavior system or the B2X system as a whole on the receiver.

187
00:20:09,280 --> 00:20:20,860
So the attacker could attack the security layer, he could attack on the L1, L2 layer, that physical MAC layer, that is, he could perform some attacks on the spectrum, or he could attack on the application layer as well.

188
00:20:20,920 --> 00:20:25,220
So for example, on security layer, what he can do, he can send outdated messages.

189
00:20:25,220 --> 00:20:31,160
That means the generation time of the message is too far behind, but you're still receiving that message now.

190
00:20:31,160 --> 00:20:36,560
So that is kind of an attack, that is kind of a message that you shouldn't have really received.

191
00:20:37,000 --> 00:20:47,660
And then the other attack is that the vehicle could just send unsigned BSMs and you just receive those messages, you say, oh, this is not authenticated, you drop that.

192
00:20:47,980 --> 00:21:08,900
Then to counter an outdated B2X message attack, the attacker could forward the generation time in the BSMs and you could keep that current to actually try to circumvent the old message check that can be performed on the security layer.

193
00:21:08,900 --> 00:21:25,080
The attacker could also send a message without a certificate to invoke no certificate check, he could also send an expired certificate, he could incorrectly sign a message, and then he could be inconsistent on the channel as well during transmission.

194
00:21:25,080 --> 00:21:34,880
Now, all of these attacks on the security layer, what they could do is they can starve the receiver of resources, of processing power.

195
00:21:34,920 --> 00:21:58,040
So even though all these messages are false and they raise all these concerns, the receiver could already lose processing any BSM that is from a legitimate vehicle and possibly not react to any real incident that might happen in front of it.

196
00:21:58,140 --> 00:22:13,300
So similarly, you look at coming to spectrum misbehavior on the physical and math layer, the attacker could, as we see in any wireless communication, could just jam the signal with the blast energy of the channel and the receiver cannot hear anything.

197
00:22:13,400 --> 00:22:15,280
Then he could also perform a DOS attack.

198
00:22:15,280 --> 00:22:17,920
For example, he can send only preambles.

199
00:22:17,920 --> 00:22:26,820
He can just say, oh, my message length is this much and does not send the message, which is a truncate after preamble attack that we have heard about recently.

200
00:22:27,220 --> 00:22:42,080
And a lot of wireless communication standards actually ask the receiver to wait for a certain period of time after the counter to wait for the message so that they can read the message during that time.

201
00:22:42,080 --> 00:22:43,820
So the attacker actually...

202
00:22:43,820 --> 00:22:49,280
So the receiver just waits and does not process any other messages coming in.

203
00:22:49,280 --> 00:22:52,520
So that can be a DOS attack that the attacker could perform.

204
00:22:53,300 --> 00:23:08,860
And then the attacker could change his MAC address so that if you filter MAC, according to, if you recognize an attacker's MAC address, if I get any message with this MAC address, I'm going to filter any messages coming from this guy.

205
00:23:08,860 --> 00:23:12,080
He can change that MAC address and still pass your filter.

206
00:23:12,240 --> 00:23:20,680
Coming to the application layer, the attacker could just fuzz all the fields, like we do in all...

207
00:23:25,460 --> 00:23:28,640
like we do in testing any mechanism.

208
00:23:28,640 --> 00:23:40,020
The attacker could just say, oh, I'm sending a speed of minus infinity to infinity and I want to see how does the way, does the receiver really accept those messages?

209
00:23:40,340 --> 00:23:47,960
Or what kind of behavior does the vehicle show me if I send such messages?

210
00:23:47,960 --> 00:23:57,120
And as Jonathan showed about civil attack, the way we could change MAC addresses, use the same certificate or different certificates and create those vehicles.

211
00:23:57,120 --> 00:24:05,340
And then lastly, the attacker could formulate specific messages to fool V2X applications.

212
00:24:05,400 --> 00:24:13,680
And we're going to talk about this in detail now with an example of the EBL application in the next couple of slides.

213
00:24:18,790 --> 00:24:31,810
So here you can see that there are two cars here, the blue one we consider as the victim vehicle, the benign vehicle, and the dark gray car is the attacker.

214
00:24:31,810 --> 00:24:39,030
Now the attacker formulates a normal BSM, which is marked all as green, where all the fields of the BSM are consistent with each other.

215
00:24:39,030 --> 00:24:41,790
You have position, you have speed, you have positive acceleration.

216
00:24:41,790 --> 00:24:45,330
You don't have any braking, you don't have any EBL flag set.

217
00:24:45,330 --> 00:24:54,950
The EBL flag is the flag that has to be set for the EBL application to actually react to a heartbreaking event.

218
00:24:54,950 --> 00:25:06,970
So whenever a vehicle, if it is hit and it is heartbreaking, the EBL flag is set and the receiver, whenever it receives the EBL application, sees that flag first and then checks other things.

219
00:25:07,230 --> 00:25:18,010
So here, the attacker at first just transfers a normal BSM, our benign vehicle receives that message and says, oh, it's a normal car, goes about on its way.

220
00:25:18,250 --> 00:25:21,830
And then now our attacker says, okay, now I'm going to start my attack.

221
00:25:21,830 --> 00:25:26,550
And now the attacker puts the EBL flag as one.

222
00:25:26,650 --> 00:25:30,710
And you can see that he has not changed all of the fields.

223
00:25:30,730 --> 00:25:34,270
So the brake is zero and the acceleration is also positive.

224
00:25:34,270 --> 00:25:36,030
This should not really be the case.

225
00:25:36,030 --> 00:25:48,470
But if you don't have any misbehavior detection, the benign vehicle, when it receives this message, it says, oh, it's a heartbreaking event, I should brake, it's safety critical to me or I might crash into the vehicle.

226
00:25:48,530 --> 00:25:52,470
And our attacker goes about his way ahead.

227
00:25:53,610 --> 00:26:04,810
So at this point, our vehicle maybe looks at, maybe he doesn't see any car in the vehicle space after a while.

228
00:26:04,810 --> 00:26:12,070
And like, oh, okay, there was no car in the vehicle space, neither physically in front of me.

229
00:26:12,070 --> 00:26:14,370
So I should just continue my motion.

230
00:26:14,370 --> 00:26:16,670
Maybe my EBL application was good.

231
00:26:16,670 --> 00:26:19,270
So this is what is the attack that we consider.

232
00:26:19,270 --> 00:26:32,270
And here we assume that the benign vehicle doesn't use his sensors like camera or radar in this example to verify messages.

233
00:26:33,330 --> 00:26:36,050
So how would you detect such kind of an attack?

234
00:26:36,050 --> 00:26:50,250
So firstly, as it might be apparent that you look at the inconsistency within the BSMPs and you say, oh, okay, so there's an inconsistency, it is possible that this message is wrong.

235
00:26:50,390 --> 00:26:53,470
And you can also check the telemetry of the attacker.

236
00:26:53,470 --> 00:27:15,370
So if the attacker, after an EBL, you would expect that the vehicle would stop, but if the vehicle continues to change its positions and has a positive speed and the vehicle keeps on moving, meaning it doesn't have any change between the telemetry that was before the attack or after the attack,

237
00:27:15,370 --> 00:27:19,150
that is another indication where how you could detect such kind of an attack.

238
00:27:19,750 --> 00:27:24,910
Now we talk about fake EBL version two attack, where our attacker has gotten smarter.

239
00:27:24,910 --> 00:27:27,730
He's like, okay, you check for both of those things.

240
00:27:27,730 --> 00:27:32,550
You check for inconsistency and you check for my telemetry.

241
00:27:32,550 --> 00:27:34,350
So I'm going to become smarter.

242
00:27:34,470 --> 00:27:39,990
Again, the attacker transmits a normal BSM, goes about his way, the benign vehicle follows it.

243
00:27:39,990 --> 00:27:42,750
Now he transmits

244
00:27:50,010 --> 00:27:57,850
a fake EBL message, formulated specifically to counter the detection in the first version.

245
00:27:57,870 --> 00:28:04,290
You can see that the EBL flag is set as one, the brake is also set as one and the acceleration is also negative.

246
00:28:04,370 --> 00:28:14,290
So when our benign vehicle receives this, his detectors don't give him any indication and this vehicle just sees, oh, okay, this is a heartbreaking event.

247
00:28:14,350 --> 00:28:18,190
And the attacker also stops in the V2X space.

248
00:28:18,190 --> 00:28:21,350
The attacker moves physically, but stops in the V2X space.

249
00:28:21,530 --> 00:28:24,170
So our benign vehicle sees, oh, there is a vehicle stopped.

250
00:28:24,170 --> 00:28:26,110
So the telemetry has changed as well.

251
00:28:26,110 --> 00:28:27,390
So I should brake.

252
00:28:27,650 --> 00:28:32,950
The victim vehicle brakes in this case and does not continue motion.

253
00:28:33,690 --> 00:28:36,210
So this is an improvement on the attack.

254
00:28:36,210 --> 00:28:38,510
Now, how would you detect this kind of an attack?

255
00:28:41,310 --> 00:28:46,170
We have looked at all the application layer side of things in the previous checks.

256
00:28:46,170 --> 00:28:51,450
Now we look at the physical layer side of things, where we look at the signal strength versus the distance.

257
00:28:51,450 --> 00:28:59,530
So as you can see on the graph on the bottom right section, on the X-axis, you have the distance and on the Y-axis, you have the signal strength.

258
00:28:59,950 --> 00:29:11,550
And what we have seen is that the signal strength should reduce as the distance increases between any two vehicles in this kind of a graph.

259
00:29:11,990 --> 00:29:27,370
But in this case, if the attacker is sending constant positions in front of you, but is actually physically changing his positions, his characteristics would look something like the red line you see.

260
00:29:27,370 --> 00:29:30,530
The distance is the same, but the signal strength keeps reducing.

261
00:29:30,530 --> 00:29:38,470
So this is how you could actually find out genuine vehicles versus an attack vehicle in such cases.

262
00:29:39,150 --> 00:29:45,730
Now our attacker thinks, okay, now you remember me and I'm going to try to circumvent this.

263
00:29:45,730 --> 00:29:51,630
So we have a version three attack that the attacker could perform as well.

264
00:29:53,190 --> 00:29:57,330
So here the attacker again, transmits a normal DSM.

265
00:29:57,630 --> 00:30:08,030
And now he again, you know, formulates a fake email message with all the fields consistent and transmit that message over our way to receive it.

266
00:30:08,030 --> 00:30:14,810
The attacker stops in vehicle space, transmits constant positions after the attack, but continues physically, right?

267
00:30:14,810 --> 00:30:17,910
So our victim vehicle again stops.

268
00:30:17,910 --> 00:30:19,890
It's just like the version two attack.

269
00:30:19,890 --> 00:30:26,550
But now what the attacker does is that the attacker goes into a silent period.

270
00:30:26,550 --> 00:30:35,190
So the attacker would go into a silent period here where it would stop transmitting any DSMs.

271
00:30:35,190 --> 00:30:50,130
So to conserve the storage space in the benign vehicle, the benign vehicle would check its database, like, oh, I have, how far, have I received any BSM recently from this vehicle, from this attack, from this attacker?

272
00:30:50,130 --> 00:30:53,490
And you would say, no, it has been a long time since I have not received it.

273
00:30:53,490 --> 00:30:54,330
Let's say five seconds.

274
00:30:54,330 --> 00:30:57,570
Five seconds is also a long time in a real-time system.

275
00:30:57,890 --> 00:31:06,070
And we assume that the benign vehicle just deletes all the information just to conserve its storage space.

276
00:31:06,350 --> 00:31:10,830
And now he doesn't see any vehicle anyways in front of it.

277
00:31:10,830 --> 00:31:14,150
So it starts continuing its motion in the V2X space.

278
00:31:14,150 --> 00:31:18,370
It doesn't see any vehicle, so it starts continuing its motion in the physical space as well.

279
00:31:18,370 --> 00:31:22,450
Now the attacker, he says, okay, now you have forgotten about me.

280
00:31:22,450 --> 00:31:28,050
He waits for this vehicle ahead, like it slows down or something, and performs the same attack.

281
00:31:28,250 --> 00:31:37,950
Now, since our benign vehicle doesn't have any history on this attacker, since he deleted this earlier, it would fall for the same attack again.

282
00:31:37,950 --> 00:31:44,630
That is what we think an improved attacker would do.

283
00:31:44,770 --> 00:31:45,690
And how would you detect?

284
00:31:45,690 --> 00:31:59,370
You would, of course, again, look at the physical properties of this kind of an attack, and possibly remember these kinds of vehicles who kind of were not, that you were not sure about, that, okay, maybe this vehicle was malicious.

285
00:31:59,370 --> 00:32:06,990
So you try to remember more, remember such kind of vehicles for a longer time to detect, to have history on them.

286
00:32:06,990 --> 00:32:21,810
So that if you have a random jump like this in the V2X space, that this gray vehicle here, the vehicle in the V2X space, jumped from back to here, you can say that this is a random jump, and this is possibly a mistake.

287
00:32:23,690 --> 00:32:48,690
Yes, so this is how we have formulated a cat and mouse game, where you come up with an attacker, like a red team, blue team kind of thing, where the attacker becomes smarter, we make our detector smarter, and we expand our detectors from the application layer to the physical layer,

288
00:32:48,690 --> 00:32:52,990
and map layer, and we expand the portfolio of all the detectors.

289
00:32:53,370 --> 00:33:02,650
And now we have a demo for this, that we would like to present there, where we try to simulate these attacks.

290
00:33:02,650 --> 00:33:07,850
And I would like to hand over the screen to my colleague, Song.

291
00:33:08,570 --> 00:33:09,590
Thank you, Rashid.

292
00:33:14,810 --> 00:33:16,570
Yeah, hello everyone.

293
00:33:16,570 --> 00:33:26,090
In the following demo, I will present the three attacking scenario as presented by Rashid, in order to show the progression of our attacks.

294
00:33:26,090 --> 00:33:30,370
So first, let's take a look at our demo setup.

295
00:33:30,370 --> 00:33:34,690
And this is a demo setup that we show in CES 2020.

296
00:33:34,690 --> 00:33:43,410
As you can see, there are two Roadrunner boards, and one HMI device, and one demo control panel.

297
00:33:43,850 --> 00:33:52,710
And a Roadrunner is a Qualcomm CV2x development platform, and that is equipped with Qualcomm CV2x chipset.

298
00:33:52,710 --> 00:33:57,090
And it is used as an onboard unit in this demo.

299
00:33:57,390 --> 00:34:08,630
Here you can see we have two Roadrunners, one of which is the genuine host vehicle, while the other one is the remote attacker vehicle.

300
00:34:09,270 --> 00:34:23,810
Since the Roadrunner can play the pre-recorded BSM files, so even though they are placed stationary, they can still transmit BSMs to each other, as if they are traveling on a real road.

301
00:34:23,810 --> 00:34:32,510
On the other hand, the HMI display is an Android application that will visualize the interaction between the two vehicles.

302
00:34:32,510 --> 00:34:35,890
And the control panel is used to control and configure our demo.

303
00:34:41,430 --> 00:34:45,730
So here is the control panel of our demo.

304
00:34:45,730 --> 00:34:54,790
And on top of the panel, you can see the option of our attack scenarios from fake EVL version one to version three.

305
00:34:54,790 --> 00:34:59,110
Actually, we have many more attack scenarios, but here we only show those three.

306
00:35:00,050 --> 00:35:12,990
And below the attack options is our misbehavior protection system status display, which will show the detection results of our misbehavior protection system.

307
00:35:13,530 --> 00:35:23,310
So now, let's start version one of the fake EVL explained by Rashid.

308
00:35:24,370 --> 00:35:37,170
In version one, the attacker generates a fake EVL warning, but keeps moving both physically as well as in the V2X space, as shown in this HMI display.

309
00:35:38,050 --> 00:35:49,270
So in this HMI display, as you can see, the green bubble here is the host vehicle, while the blue bubble is the remote vehicle.

310
00:35:49,270 --> 00:35:55,070
From the perspective of the host vehicle, based on the BSM it received.

311
00:35:55,070 --> 00:36:01,090
And the overlay here on the right of the display is not a vehicle.

312
00:36:01,090 --> 00:36:09,570
Actually, it's a dashboard for the host vehicle, which will indicate the safety warnings generated by the safety applications.

313
00:36:10,090 --> 00:36:12,890
Let's resume the demo.

314
00:36:12,990 --> 00:36:25,550
So as you can see, from the perspective of the host vehicle, both are moving normally until the host vehicle received a fake EVL warning.

315
00:36:25,610 --> 00:36:37,910
And this is because, just as Rashid explained, because of the inconsistency between the position field and the EVL flag in the BSMs.

316
00:36:37,910 --> 00:36:43,530
So you can see the blue bubble is still moving, but it transmits a fake EVL warning.

317
00:36:44,250 --> 00:36:54,010
However, our misbehavior protection system can easily detect this inconsistency by applying the possibility check and the consistency check.

318
00:36:54,370 --> 00:36:58,110
So this is a very easy attacking scenario.

319
00:36:58,110 --> 00:37:03,610
Of course, a smart attacker will come up with smarter attacks.

320
00:37:04,530 --> 00:37:12,350
Now let's look at the second one, the version two of our fake EVL attack.

321
00:37:12,350 --> 00:37:32,690
So in version two, explained by Rashid, the attacker will generate a fake EVL warning, but all the fields in the BSM are consistent, meaning that the attacker will move physically, but it will stop in the V2X space.

322
00:37:35,150 --> 00:37:36,830
Now let's switch into the...

323
00:37:37,350 --> 00:37:49,550
So as you can see, in this version, the blue bubble is moving normally in the beginning, but then it will stop in the V2X.

324
00:37:50,270 --> 00:37:56,150
As you can see here, the blue bubble indeed stopped in the V2X space.

325
00:37:56,190 --> 00:38:01,710
So the host vehicle, we believe, is a normal vehicle in the application layer.

326
00:38:01,790 --> 00:38:10,070
However, if we're applying our physical layer detector, such as the signal strength, we can detect such misbehavior.

327
00:38:10,930 --> 00:38:16,510
So our host vehicle can ignore this fake EVL warning and continue the motion.

328
00:38:17,130 --> 00:38:24,350
You can see here, our host vehicle is still moving, even though it detects this, because it detects this misbehavior.

329
00:38:26,350 --> 00:38:32,550
The attacker can keep finding ways to upgrade his attack, explained by Rashid.

330
00:38:32,550 --> 00:38:44,670
And so in version three, and the attacker, the host vehicle, the attacker will enter a silent period after performing the first fake EVL attack.

331
00:38:44,890 --> 00:38:54,230
And then he will disappear from the V2X space, hoping that the host vehicle will forget about him and then he will show up again.

332
00:38:54,350 --> 00:38:56,650
Now let's take a look at this scenario.

333
00:38:56,650 --> 00:39:01,150
So in the beginning, the blue bubble behaves normally.

334
00:39:02,010 --> 00:39:04,650
Then he will perform the fake EVL version two.

335
00:39:04,650 --> 00:39:13,090
As you can see, in the V2X space, but then it disappeared, hoping the host vehicle will forget about him.

336
00:39:13,090 --> 00:39:18,990
Now, after the silent period, it shows up again and then perform the attack again.

337
00:39:19,910 --> 00:39:27,870
So again, this attack can also be detected using the physical layer detectors just introduced by Rashid.

338
00:39:32,990 --> 00:39:33,890
Okay.

339
00:39:34,390 --> 00:39:45,610
So as the demo shows that a smart attacker can always find ways to create smarter attacks as our detection algorithm evolves.

340
00:39:45,610 --> 00:39:54,630
For example, we showed that using the physical layer, we can detect the fake EVL attack version two and version three.

341
00:39:54,630 --> 00:40:04,730
However, a smart attacker may circumvent our detector by, for example, matching the signal strengths with valid distribution.

342
00:40:04,730 --> 00:40:10,130
Valid signal strength distribution by adjusting the transmission power.

343
00:40:10,270 --> 00:40:22,970
And other cases, attacker can use different MAC address and pseudonym certificates to perform a stable attack as introduced previously to fool the majority based detector.

344
00:40:22,970 --> 00:40:31,050
And actually there are many more ways that attacker can progress to perform more advanced attacks.

345
00:40:31,050 --> 00:40:42,730
However, by improving our detection system, we are able to squeeze the attacker to the corner and raise the bar and the cost of the successful attack.

346
00:40:42,730 --> 00:40:49,590
And on the other hand, the progression of the attack can push for a stronger and more robust detection system.

347
00:40:50,390 --> 00:40:51,730
Thank you.

348
00:40:59,800 --> 00:41:01,020
Thank you, Rashid.

349
00:41:01,020 --> 00:41:05,680
So we are coming to the conclusion of this talk.

350
00:41:05,680 --> 00:41:16,360
So what we showed you today is why, what is misbehavior, first of all, in V2X and give you many examples.

351
00:41:16,540 --> 00:41:23,540
And then we even went through this misbehavior life cycle, detection, local reaction, reporting, investigation, global reaction.

352
00:41:23,820 --> 00:41:27,500
And then we demonstrated the progression of an attack.

353
00:41:27,500 --> 00:41:30,220
What does the attacker do to be smarter and smarter?

354
00:41:30,220 --> 00:41:32,840
What would you test when you want to attack the system?

355
00:41:33,300 --> 00:41:36,880
So, and we'll give you a demo of this working on real onboard unit.

356
00:41:37,140 --> 00:41:46,080
So as a conclusion, really we understand why misbehavior protection is super important for more secure V2X deployment.

357
00:41:46,080 --> 00:41:49,360
You need to have it, otherwise authentication isn't sufficient.

358
00:41:49,360 --> 00:41:50,440
You will see this big data.

359
00:41:50,440 --> 00:41:52,980
You need to be able to analyze it and detect it.

360
00:41:52,980 --> 00:41:59,100
But what's good is that misbehavior detection, reporting, investigation, reaction are all being standardized right now.

361
00:41:59,100 --> 00:42:08,140
So we hope that then really we're going to raise that bar on making it harder for an attacker and all have at least minimum performance in terms of detection.

362
00:42:08,280 --> 00:42:17,740
And also what is really positive is that misbehavior protection solutions exist and have been deployed in, for example, one in the connected vehicle part of deployment.

363
00:42:17,860 --> 00:42:23,140
So this is positive, you know, it's there, we are ahead of the curve, basically.

364
00:42:23,380 --> 00:42:29,820
But as we've seen, the attacker is always smarter and will try to counter what you do.

365
00:42:29,820 --> 00:42:34,940
So we still have a lot of open challenges and I really just like to give you three examples here.

366
00:42:34,960 --> 00:42:46,800
The first one is, if you are using a machine learning based detector, therefore an attacker will use a visual example, a visual machine learning to fool your detection algorithm.

367
00:42:46,800 --> 00:42:51,740
So this is a really active field right now to look at visual machine learning.

368
00:42:51,860 --> 00:42:54,500
And so therefore, if you're based on that, you need to look at it.

369
00:42:55,760 --> 00:43:02,420
This is still a highly sophisticated level of attack, but it is a matter of time before it's making it easier to be there.

370
00:43:02,460 --> 00:43:07,660
The second part is that we've shown today only the detection, the attacks on the first step.

371
00:43:07,660 --> 00:43:12,160
So you need to investigate the attack on the reporting, investigation and a global reaction.

372
00:43:12,200 --> 00:43:14,240
What could you do on those different stages?

373
00:43:14,840 --> 00:43:23,040
And finally, don't forget that V2X will be part of an automated vehicle and therefore will be fused with other sensors.

374
00:43:23,040 --> 00:43:33,060
So you need to analyze the effect of V2X attack on the sensor fusion to understand how much inconsistency and insensitivity can you create just from the V2X standpoint.

375
00:43:33,520 --> 00:43:38,140
With that, this is the end of our presentation and we'll be happy to take questions.

376
00:43:38,140 --> 00:43:39,020
Thank you.