3 ms·

People can enjoy the benefits of AI assistance without having their privacy violated. Running ML models on edge devices will become the norm whenever possible.

People can enjoy the benefits of AI assistance without having their privacy violated. Running ML models on edge devices will become the norm whenever possible. Then we will allow such devices like Alexa to talk to web endpoints to get data, but not to submit any info the other way. I definitely can see a whole new niche of software called "privacy protection firewalls" that would sit on the network between "Alexa" and your wifi router (perhaps running on the router) and do smart filtering what "Alexa" and other devices alike are allowed to do. We're not very far from being able to run pretty powerful AIs on edge devices. For one today's newest android phones all support vulkan and are equipped with GPUs that can run pretty sophisticated models. Also our better understanding of those models will lead to optimisations in their size. So, let's not throw the baby out with the bath water. Let's not paint AI assistants watching over out sleep as something bad. It is not. It is bad only when we have no control over what it does with the data it collects.

bo1024 4y ago
Agree with a lot of this, but what a sad situation if the best we can hope for in "control over what it does" is distrust my own device so much that I install firewalls to keep it in check. A much better version of this vision is FOSS.
- notriddle 4y ago
  FOSS is useful, and may be a part of the solution, but it is not the solution in and of itself the way Free Software advocates portray it as. Most people aren't programmers, people who are programmers don't read the source code of everything they run even if they're allowed to, and even people who try to read the source code for everything they run won't necessarily be able to catch every potential nefarious act. The ingredients list on the back of the box is not the only form of consumer protection in place for food. It shouldn't be the only consumer protection in place for software, either.
mpalmer 4y ago
That's interesting. It's like the "edge" in edge computing/caching is sort of a fractal developing over time. An edge becoming more and more tightly looped and furled, while topologically, the network stays the same shape. It's market forces that are causing the edge to fractalize, pushing the computing and storage hardware closer to the customer, but only to a point. It may take a different market and different forces to push the edge to its end state. I think centralized compute and storage could very well end up ceding some ground to private, redundant, locally networked dedicated computing/storage substrates, which might take increasingly smaller and more exotic forms. What an interesting time that would be!
wink 4y ago
Sounds like a smart and easy solution, but how do you think this will have any remote chance of being doable? Your gadget communicates with the outside world, let's assume HTTP+TLS. Ok, so you need to MITM it. Does it speak plaintext? Ok, how do you know a=1234234&b=30487234807 does not say anything about your habits? Does it speak a binary protocol? Now you need to reverse that. No. Either I trust the device or I don't.
- Roark66 4y ago
  Both issues you raise are already kind of solved(or very close to it) in the corporate security world. Also, I think people will lie and cheat when we make it easy for them to get away with it. Sending all data their way into their closed "black box systems" and trusting they don't use it for nefarious things makes is extremely easy for them to do whatever they want. Having devices run mostly at the edge communicating sporadically with the mothership or other web services it takes much more effort to conceal private information gathering and many companies will think it is not worth the risk of getting caught Also, no matter how cleverly they conceal it, someone will be able to reverse engineer it. When all we send is camera/microphone/sensor data this is 100% not possible. About your specific points. Yes, MiTM would be required today although under a nicer name. In "corporate security" context this is often called "SSL inspection" and various hacks to end devices are required to make it work well. In future I hope no hacks will be required for home devices and enabling it will be a matter of a simple setting. Then a router/appliance could be used as SSL traffic decrypting proxy. One way it could be done, a customer could push a button on a device for it to download a self signed certificate from the router/scanning appliance. Then it would trust that cert in place of the target while the router would handle the ssl between itself and the remote service. Then you raise a point of traffic scanning. It is not a trivial problem, but much of the technology already exists in various (again corpo) security scanning products. From signature based, through heuristics, to perhaps future AI based methods it is possible to achieve some level of protection from information leakage if you are able to scan all outgoing traffic and devices that communicate are generally built supporting well defined protocols. When AI runs at the edge at least we have a fighting chance to succeed most of the time. When it runs in the cloud and we just stream camera/microphone/sensor data there is no hope.

bo1024 4y ago

Agree with a lot of this, but what a sad situation if the best we can hope for in "control over what it does" is distrust my own device so much that I install firewalls to keep it in check. A much better version of this vision is FOSS.

notriddle 4y ago
FOSS is useful, and may be a part of the solution, but it is not the solution in and of itself the way Free Software advocates portray it as. Most people aren't programmers, people who are programmers don't read the source code of everything they run even if they're allowed to, and even people who try to read the source code for everything they run won't necessarily be able to catch every potential nefarious act. The ingredients list on the back of the box is not the only form of consumer protection in place for food. It shouldn't be the only consumer protection in place for software, either.

mpalmer 4y ago

That's interesting. It's like the "edge" in edge computing/caching is sort of a fractal developing over time. An edge becoming more and more tightly looped and furled, while topologically, the network stays the same shape. It's market forces that are causing the edge to fractalize, pushing the computing and storage hardware closer to the customer, but only to a point. It may take a different market and different forces to push the edge to its end state. I think centralized compute and storage could very well end up ceding some ground to private, redundant, locally networked dedicated computing/storage substrates, which might take increasingly smaller and more exotic forms. What an interesting time that would be!

wink 4y ago

Sounds like a smart and easy solution, but how do you think this will have any remote chance of being doable? Your gadget communicates with the outside world, let's assume HTTP+TLS. Ok, so you need to MITM it. Does it speak plaintext? Ok, how do you know a=1234234&b=30487234807 does not say anything about your habits? Does it speak a binary protocol? Now you need to reverse that. No. Either I trust the device or I don't.

Roark66 4y ago
Both issues you raise are already kind of solved(or very close to it) in the corporate security world. Also, I think people will lie and cheat when we make it easy for them to get away with it. Sending all data their way into their closed "black box systems" and trusting they don't use it for nefarious things makes is extremely easy for them to do whatever they want. Having devices run mostly at the edge communicating sporadically with the mothership or other web services it takes much more effort to conceal private information gathering and many companies will think it is not worth the risk of getting caught Also, no matter how cleverly they conceal it, someone will be able to reverse engineer it. When all we send is camera/microphone/sensor data this is 100% not possible. About your specific points. Yes, MiTM would be required today although under a nicer name. In "corporate security" context this is often called "SSL inspection" and various hacks to end devices are required to make it work well. In future I hope no hacks will be required for home devices and enabling it will be a matter of a simple setting. Then a router/appliance could be used as SSL traffic decrypting proxy. One way it could be done, a customer could push a button on a device for it to download a self signed certificate from the router/scanning appliance. Then it would trust that cert in place of the target while the router would handle the ssl between itself and the remote service. Then you raise a point of traffic scanning. It is not a trivial problem, but much of the technology already exists in various (again corpo) security scanning products. From signature based, through heuristics, to perhaps future AI based methods it is possible to achieve some level of protection from information leakage if you are able to scan all outgoing traffic and devices that communicate are generally built supporting well defined protocols. When AI runs at the edge at least we have a fighting chance to succeed most of the time. When it runs in the cloud and we just stream camera/microphone/sensor data there is no hope.