|
boner confessor posted:this is potentially more dangerous because imagine driving a car which mostly drives itself but you still have to pay attention, not doing anything too distracting, ready to take over when the computer says "i dont know what to do, returning to manual mode in 3... 2... 1..." how many people are going to be shutting off porn and getting their dicks caught in the zipper? how many people are going to be watching tv or reading a book?
|
#
?
Aug 23, 2017 22:02
|
|
|
|
| # ? Jun 9, 2024 03:53 |
|
|
Condiv posted:cause if not, a street lamp with a sticker on it that tells driverless cars "school zone ended, speed limit 150" and is not recognizable to humans as any kind of attack is just as effective as using these attacks on stopsigns this kicked off a big derail in the past but imo self driving cars are going to default to a stored database of road data for routing and basic safety info like school zone, speed limits etc. in the absence of potentially unreliable signage data Cicero posted:Depends on what the 5% is. If it's "you can only self-drive in regions we've extensively mapped" then it's not really a problem. sudden hazardous weather most likely, or detours around temporary lane closures like for accidents - ideally neither being approached at speed
|
|
#
?
Aug 23, 2017 22:03
|
|
|
boner confessor posted:this kicked off a big derail in the past but imo self driving cars are going to default to a stored database of road data for routing and basic safety info like school zone, speed limits etc. in the absence of potentially unreliable signage data you can't have a realistically self-driving system that relies on stored databases, cause stored databases != real life, current conditions. driverless cars need to deal with construction, or police rerouting cause of an accident, or or or. worse, stored databases are also an easy target if you want to cause some mayhem. it's not like hackers have failed to break into a database before, and loving up a routing and basic safety info database would cause a whole lot of mayhem and destruction until it was fixed. that's not to mention that adversarial attacks don't have to be based on signage or image recognition to be effective. i'm sure you could abuse a car's lidar input to create obstacles where there are none to interfere with traffic and all kinds of other nasty attacks that we haven't thought of yet. Condiv fucked around with this message at 22:22 on Aug 23, 2017 |
|
#
?
Aug 23, 2017 22:20
|
|
|
Did I dream up the self driving car thread someone made to quarantine this poo poo? If I did then please, someone who cares, make it.
|
|
#
?
Aug 23, 2017 22:22
|
|
|
I can't tell if the point of better technology is to improve our lives or to avoid everyone else by never having to leave the house.
|
|
#
?
Aug 23, 2017 22:38
|
|
|
Star Man posted:I can't tell if the point of better technology is to improve our lives or to avoid everyone else by never having to leave the house. Seems like you're saying the same thing here, friendo. 
|
|
#
?
Aug 23, 2017 22:39
|
|
|
Lightning Lord posted:Seems like you're saying the same thing here, friendo. I considered that, yes.
|
|
#
?
Aug 23, 2017 22:40
|
|
|
Condiv posted:you can't have a realistically self-driving system that relies on stored databases, cause stored databases != real life, current conditions. driverless cars need to deal with construction, or police rerouting cause of an accident, or or or. worse, stored databases are also an easy target if you want to cause some mayhem. it's not like hackers have failed to break into a database before, and loving up a routing and basic safety info database would cause a whole lot of mayhem and destruction until it was fixed. that's not to mention that adversarial attacks don't have to be based on signage or image recognition to be effective. i'm sure you could abuse a car's lidar input to create obstacles where there are none to interfere with traffic and all kinds of other nasty attacks that we haven't thought of yet. yeah it's a supplement friendo. that's why i said "default to". as i recall the last derail came because people cant read and assumed i was talking about doing 100% sql lookups for basic self driving tasks even though if you just think about it for half a second that's so implausible it couldn't possibly be what i am saying road conditions and networks don't change often enough to make stored databases unrealistic for the majority of real world needs, this is already how GPS navigation systems work either as a standalone device in your car or as data streamed to you from somewhere. they can even notify you when you're driving over the referenced speed limit on a given segment of road
|
|
#
?
Aug 23, 2017 22:43
|
|
|
Star Man posted:I can't tell if the point of better technology is to improve our lives or to avoid everyone else by never having to leave the house. You're probably right though that this will enable more people to live as hikikomori.
|
|
#
?
Aug 23, 2017 22:53
|
|
|
boner confessor posted:yeah it's a supplement friendo. that's why i said "default to". as i recall the last derail came because people cant read and assumed i was talking about doing 100% sql lookups for basic self driving tasks even though if you just think about it for half a second that's so implausible it couldn't possibly be what i am saying you're either misusing "default to" or misstating how your proposed system would work. you're saying it would look up the database whenever it could and rely on visual and topological data second if the database is unavailable, which is a disaster waiting to happen for the reasons i already told you. if you mean it would look up database info and then use that to supplement real world info collected by the lidar and other sensors, well, you've just hit the adversarial data problem again and that database lookup isn't going to protect you from it. i would suggest you rethink what you're saying cause it doesn't make sense that you think the car wouldn't be hitting the database a ton, but also would default to the database. quote:road conditions and networks don't change often enough to make stored databases unrealistic for the majority of real world needs, this is already how GPS navigation systems work either as a standalone device in your car or as data streamed to you from somewhere. they can even notify you when you're driving over the referenced speed limit on a given segment of road those real world needs currently don't involve guiding a car, and GPS is not sufficient for that or we would already have self driving cars. you're oversimplifying the problem and then saying it's easily solvable. Condiv fucked around with this message at 22:59 on Aug 23, 2017 |
|
#
?
Aug 23, 2017 22:56
|
|
|
Condiv posted:default to is not a supplement in the traditional sense. you're saying it would look up the database whenever it could and rely on visual and topological data second if the database is unavailable, which is a disaster waiting to happen for the reasons i already told you. i would suggest you rethink what you're saying cause it doesn't make sense that you think the car wouldn't be hitting the database a ton, but also would default to the database. yeah it's way more likely you're not actually paying attention to what i'm saying at all and are just making up some argument to respond to in your own imagination. people already use database-driven GPS navigation and have for years, if you think it's suddenly implausible you're just being a contrarian or you simply are not paying attention to my actual argument and i've lost any interest in trying to get over whatever communicative roadblocks you have e: haha you're probably thinking i'm saying LIDAR is unnecessary because of GPS, and if so then yeah you're just hung up on whatever it is you're thinking about and you're not willing or capable of reading my actual argument. goons are generally bad at reading comprehension but jeez VVV lol yeah sure i'll do that, guy ironically arguing about sensing bad information leading to an unfortunate outcome who can't slow down and read a post to understand how to formulate a valid response boner confessor fucked around with this message at 23:10 on Aug 23, 2017 |
|
#
?
Aug 23, 2017 22:58
|
|
|
boner confessor posted:yeah it's way more likely you're not actually paying attention to what i'm saying at all and are just making up some argument to respond to in your own imagination. people already use database-driven GPS navigation and have for years, if you think it's suddenly implausible you're just being a contrarian or you simply are not paying attention to my actual argument and i've lost any interest in trying to get over whatever communicative roadblocks you have a self-driving car is different than people. people can use GPS cause they use it to supplement their senses. a self driving car with a neural network cannot trust its senses in the face of adversarial data, and GPS data will not protect it from that (cause at that point you're choosing to ignore potential real world issues in lieu of a database that may be out of date!) once you've introduced lidar, or other sensors, you hit the adversarial data problem and you have to choose to either trust the neural network of the car or the database info, and you'd be insane to chose the database info rethink your argument cause it's idiotic and doesn't actually apply to the conversation
|
|
#
?
Aug 23, 2017 23:01
|
|
|
Default can be used in either sense, stop slap fighting over the definition. Clearly what was meant was that a car would fall back to stored data in the absence of good sensor data. Stored maps can be updated in seconds with LTE connected cars, I literally worked on software to do that.
|
|
#
?
Aug 23, 2017 23:07
|
|
|
baquerd posted:Default can be used in either sense, stop slap fighting over the definition. Clearly what was meant was that a car would fall back to stored data in the absence of good sensor data. Stored maps can be updated in seconds with LTE connected cars, I literally worked on software to do that. but it's not reliable! what fool would ever trust google maps?
|
|
#
?
Aug 23, 2017 23:08
|
|
|
baquerd posted:Default can be used in either sense, stop slap fighting over the definition. Clearly what was meant was that a car would fall back to stored data in the absence of good sensor data. Stored maps can be updated in seconds with LTE connected cars, I literally worked on software to do that. and that still doesn't help cause if detecting if data was "good" or not was feasible, adversarial data would not be a problem. computer scientists have been trying to find ways to detect if data is good or not to shield DNNs from adversarial attacks and none of their efforts have panned out yet. sorry, but you guys have not found the one weird trick to fixing this problem
|
|
#
?
Aug 23, 2017 23:09
|
|
|
baquerd posted:Default can be used in either sense, stop slap fighting over the definition. Clearly what was meant was that a car would fall back to stored data in the absence of good sensor data. Stored maps can be updated in seconds with LTE connected cars, I literally worked on software to do that. So who will be responsible for mapping rerouted traffic due to construction?
|
|
#
?
Aug 23, 2017 23:09
|
|
|
i have specifically already told you guys that adversarial data cannot be detected by a neural network as faulty, that's why it's as insidious as it is. that you guys think: "oh, the self-driving car will see the adversarial data, realize it's bad and look up the database" shows that you're not listening at all. if determining "this data is bad!!" was easy (or even actually doable at the moment) then adversarial data wouldn't be an issue and could be guarded against!
|
|
#
?
Aug 23, 2017 23:12
|
|
|
hobbesmaster posted:So who will be responsible for mapping rerouted traffic due to construction? Basically, the cars themselves will communicate sensor data to the cloud for combined processing and distribution of the new lane definitions. There are POC swarming/follow the leader algorithms in use for emergencies/towing use cases that might be helpful. Cars can default in risky situations to slower speeds, increasing analysis time.
|
|
#
?
Aug 23, 2017 23:13
|
|
|
hobbesmaster posted:So who will be responsible for mapping rerouted traffic due to construction? once you have a database of roads and an algorithm to calculate a route, all you have to do is calculate a new route which does not use the road segment under construction. right now there's no national standard for communicating road network data to routing database owners like this but there probably will be in the future so the state DOT can not only push notifications (which google can already collect via user input) along with preferred detours Condiv posted:i have specifically already told you guys that adversarial data cannot be detected by a neural network as faulty, that's why it's as insidious as it is. ah yes i remember the epidemic levels of google maps and garmin/navteq data being man in the middled and people being maliciously routed all over the place shut up nerd
|
|
#
?
Aug 23, 2017 23:13
|
|
|
baquerd posted:Basically, the cars themselves will communicate sensor data to the cloud for combined processing and distribution of the new lane definitions. There are POC swarming/follow the leader algorithms in use for emergencies/towing use cases that might be helpful. Cars can default in risky situations to slower speeds, increasing analysis time. how does the car determine if it's in a risky situation in the face of data that is specifically designed to trick it to think otherwise? again, if adversarial data was easy to recognize it wouldn't be a problem
|
|
#
?
Aug 23, 2017 23:14
|
|
|
boner confessor posted:once you have a database of roads and an algorithm to calculate a route, all you have to do is calculate a new route which does not use the road segment under construction. right now there's no national standard for communicating road network data to routing database owners like this but there probably will be in the future so the state DOT can not only push notifications (which google can already collect via user input) along with preferred detours That doesn't help with the typical interstate construction though. Let's not even get started with weather...
|
|
#
?
Aug 23, 2017 23:16
|
|
|
Condiv posted:i have specifically already told you guys that adversarial data cannot be detected by a neural network as faulty, that's why it's as insidious as it is. If there is a big discrepancy between map and sensor data, we already have algorithms to limit that impact. For example, in Germany they have variable speed signs which change speed limits, sometimes unpredictably, and car image recognition isn't 100%. Continual large discrepancies would get flagged and investigated by an employee or government official.
|
|
#
?
Aug 23, 2017 23:16
|
|
|
hobbesmaster posted:That doesn't help with the typical interstate construction though. if an interstate isn't entirely closed then you don't need to worry about alternate routing or detours, you just spend more time sitting in traffic which isn't so bad since you're in a self driving car now also there's a difference here between how self driving cars will navigate their local environment, i.e. maintain lane and not hit obstacles, versus how they will navigate the overall road network, i.e. get from A to B. the tricky part is not hitting obstacles, because the other part has been solved for years now. weather doesn't have much to do with the latter part, routing, because you're going to take the same route whether or not it's raining (unless your route has flooded or something)
|
|
#
?
Aug 23, 2017 23:18
|
|
|
baquerd posted:If there is a big discrepancy between map and sensor data, we already have algorithms to limit that impact. For example, in Germany they have variable speed signs which change speed limits, sometimes unpredictably, and car image recognition isn't 100%. Continual large discrepancies would get flagged and investigated by an employee or government official. or, you know, the people getting into wrecks or their reporting their car malfunctioning. at which point it's already too late, the damage has been done. what you're describing is not a particularly effective defense against adversarial data considering how easy it is to make and put into use
|
|
#
?
Aug 23, 2017 23:19
|
|
|
hobbesmaster posted:That doesn't help with the typical interstate construction though. Yeah, weather is a bitch. Interstate construction that just moves the lanes away from the paints lines is pretty well solved though.
|
|
#
?
Aug 23, 2017 23:20
|
|
|
Condiv posted:or, you know, the people getting into wrecks or their reporting their car malfunctioning. at which point it's already too late, the damage has been done. There is rather more rudimentary technology in place to simply prevent the cars from crashing though, as had been mentioned.
|
|
#
?
Aug 23, 2017 23:21
|
|
|
baquerd posted:There is rather more rudimentary technology in place to simply prevent the cars from crashing though, as had been mentioned. what? you have mentioned nothing that actually prevents these attacks and I have repeatedly pointed out to you why that is. slapping a GPS map onto a system doesn't fix the issue
|
|
#
?
Aug 23, 2017 23:23
|
|
|
baquerd posted:Yeah, weather is a bitch. Interstate construction that just moves the lanes away from the paints lines is pretty well solved though. Not implemented in mapping software at least. "Proceed to the route" over and over again when you're routed on the other side of the highway for example.
|
|
#
?
Aug 23, 2017 23:23
|
|
|
Do Self-Driving Cars Dream Of Electric Unicorns? The latest
|
|
#
?
Aug 23, 2017 23:24
|
|
|
hobbesmaster posted:Not implemented in mapping software at least. well yeah, you don't need mapping or navigational software to tell you how to continue going down a road past some construction signs. that's part of the basic "don't crash into anything' software, not "how do i get to my destination"
|
|
#
?
Aug 23, 2017 23:26
|
|
|
Condiv posted:what? you have mentioned nothing that actually prevents these attacks and I have repeatedly pointed out to you why that is. slapping a GPS map onto a system doesn't fix the issue These attacks only target one part of a much greater system. It will take a great deal of ingenuity to actually get a self driving car in dangerous situations because telling it that the speed limit is 150 mph or that it's going the wrong way down a one way street already happens! Image recognition is not relied on as anything but one more piece of potentially faulty data.
|
|
#
?
Aug 23, 2017 23:27
|
|
|
It feels like you guys are reiterating the same arguments that happened in the automation thread couple days ago. Mostly automated driving of any vehicle land, sea, air, or space, is quite possible now, but this involves a human intelligence acting as a tie breaker whenever competing information exists or information is absent. You can duplicate sensors, data, all sorts of things, but in the end it still comes back to a single decision making centre choosing whether to stop, go, turn, etc. Until that single decision making centre is intelligent enough to not only use data, but assess its quality, it's difficult to call it capable of unattended function.
|
|
#
?
Aug 23, 2017 23:27
|
|
|
hobbesmaster posted:Not implemented in mapping software at least. "Proceed to the route" over and over again when you're routed on the other side of the highway for example. It's being poc'd with major German auto makers. Some forms of this will be built into next years luxury models.
|
|
#
?
Aug 23, 2017 23:28
|
|
|
baquerd posted:These attacks only target one part of a much greater system. It will take a great deal of ingenuity to actually get a self driving car in dangerous situations because telling it that the speed limit is 150 mph or that it's going the wrong way down a one way street already happens! Image recognition is not relied on as anything but one more piece of potentially faulty data. and i have already posted in this thread that adversarial data is not specific to image recognition or convolutional neural networks! the technique exploits flaws in the way sigmoid neurons work, which are the building blocks of current day neural networks. you can create adversarial sounds to fool voice recognizing neural nets, adversarial games to fool and trip up go-playing neural networks, and adversarial topological data to fool neural networks relying on lidar.
|
|
#
?
Aug 23, 2017 23:32
|
|
|
Maluco Marinero posted:It feels like you guys are reiterating the same arguments that happened in the automation thread couple days ago. Mostly automated driving of any vehicle land, sea, air, or space, is quite possible now, but this involves a human intelligence acting as a tie breaker whenever competing information exists or information is absent. You can duplicate sensors, data, all sorts of things, but in the end it still comes back to a single decision making centre choosing whether to stop, go, turn, etc. yeah pretty much this
|
|
#
?
Aug 23, 2017 23:33
|
|
|
Condiv is getting ahead of himself about the success of adversarial data - there's never been a trial of this sort of data against autonomous vehicle systems. So far, they've fooled sign classifiers in lab conditions, e.g. Evtimov et al. from this year. That's a pretty serious step from sign misidentification to aberrant car behavior.
|
|
#
?
Aug 23, 2017 23:33
|
|
|
in 90% of situations anyway a self driving car is just going to do what humans do, which is do what the other cars are doing. works real well so long as you're not at the front, and you dont need too many fancy tricks to maintain X distance from the vehicle ahead of you staying below Y speed
|
|
#
?
Aug 23, 2017 23:35
|
|
|
a foolish pianist posted:Condiv is getting ahead of himself about the success of adversarial data - there's never been a trial of this sort of data against autonomous vehicle systems. So far, they've fooled sign classifiers in lab conditions, e.g. Evtimov et al. from this year. That's a pretty serious step from sign misidentification to aberrant car behavior. i am not, but you can pretend i am if you wish. there is not anything that specially seperates autonomous vehicle systems from other state of the art DNNs for the purposes of these attacks. they are still vulnerable to these attacks because of their reliance on sigmoid neurons for learning.
|
|
#
?
Aug 23, 2017 23:36
|
|
|
Condiv posted:and i have already posted in this thread that adversarial data is not specific to image recognition or convolutional neural networks! the technique exploits flaws in the way sigmoid neurons work, which are the building blocks of current day neural networks. you can create adversarial sounds to fool voice recognizing neural nets, adversarial games to fool and trip up go-playing neural networks, and adversarial topological data to fool neural networks relying on lidar. And there are data quality and verifiability metrics. Keep in mind that these attacks are almost always demonstrated against naive systems.
|
|
#
?
Aug 23, 2017 23:36
|
|
|
|
| # ? Jun 9, 2024 03:53 |
|
|
Condiv posted:i am not, but you can pretend i am if you wish. there is not anything that specially seperates autonomous vehicle systems from other state of the art DNNs, which are still vulnerable to these attacks because of their reliance on sigmoid neurons for learning. Seriously, just read that paper. It's not nearly as clear a danger as you're stating.
|
|
#
?
Aug 23, 2017 23:37
|
|