Artificial Intelligence In Every Device Goal Aided By Blockchain’s Privacy Featuresbr>
Dr. Rand Hindi is the CEO and co-founder of Snips, a platform boasting 14,000 developers who have built 20,000+ AI voice assistants, making it one of the largest voice assistant developer communities outside of Google and Amazon.
They’re now getting into blockchain to keep AI assistants private, using decentralized machine learning, a token-curated app store, and decentralized data generation. The goal is to offer a truly private-by-design alternative to the existing centralized AI assistants, such as Google Home, Alexa etc. The idea is to avoid situations like when one family’s Echo sent a private conversation to a random contact.
Over a four-year period, the Snips team researched and developed technology to use Edge computing so that user data is processed 100% locally, on-device. Now, with blockchain, they are looking to take privacy and fairness for AI assistants to the next level.
Rand Hindi talked with Block Tribune about the venture.
BLOCK TRIBUNE: What went wrong in that notorious incident where Echo sent a private conversation out?
RAND HINDI: That particular incident was due to compounding machine learning failures: Alexa thought it was being talked to, then thought it needed to send a message, then misheard the name of someone. Unfortunately, this is hardly preventable as it is the nature of machine learning: nothing is 100% accurate!
BLOCK TRIBUNE: What are the privacy risks related to Alexa?
RAND HINDI: Risks are pretty significant both for individuals and for companies. Our voice is very sensitive data. It turns out (surprise!) that nobody likes to be heard in his house. Even in the United States, which is more laxist on these issues, more than 48% of consumers are worried about how smart speakers are processing their data. In France, the CNIL recently published note denouncing the risks of these objects and recommended turning them off when they are not used!
By cloud technologies to process voice these companies simply put at risk their independence. In addition to losing control of their brands by forcing their users to trigger voice assistants saying “Alexa” or “OK Google”, they share all their user data, whether it’s what they say or what they do. These easily become satellites of big tech giants, which has a significant impact on their brand independence and sovereignty.
BLOCK TRIBUNE: How can blockchain prevent this?
RAND HINDI: One of the ways how we can solve this problem is decentralization.
Existing voice assistants represent everything that’s wrong with the internet: they use open source software but don’t contribute. They centralize all our personal data in one place, making us all vulnerable to mass surveillance and hacking. They exploit and monetize our personal data without giving us anything in exchange. They steal our privacy and the one of our families. They exploit developers without caring about their future, asking them to publish their products on app stores before arbitrarily kicking them out.
Decentralizing voice assistants is how we can fix that. By processing data on device, we don’t need to send it to the cloud. By using a blockchain to decentralize machine learning, we can reward users for their data, without ever actually access it. By using a token to power an app store, we can let the community decide what can be built and by whom. Basically, we would know nothing about its users, nothing about their lives and have no say in what people can do with their assistant.
BLOCK TRIBUNE: What is decentralized data generation and how will it help enhance privacy and fairness?
RAND HINDI: Our goal at Snips is to provide the first private-by-design voice assistant. To achieve this, we combine two technologies built at Snips: on-device processing and decentralized encrypted machine learning.
On-device processing means that when a user speaks to its Snips device, the device itself does the processing, without any data being sent to the cloud. This means nobody can listen or access their data, not even us! This “zero-data” policy guarantees Privacy by Design.
But of course, the assistant needs to get better over time, as more people use it. To achieve this without any impact of privacy, we created a new technology that leverages federated learning, multi-party computation, secret sharing and a blockchain to train a neural network without anyone ever having access to the actual user data.
This is how it works:
- Decentralized data generation, where developers can crowdsource “fake” user data from people in our community, paying them via the token. This enables developers to train their apps with a lot of data even before launching it, solving the cold start problem.
- Encrypted federate learning, where developers can improve the quality of their apps as people use them, by using only encrypted data, without any compromise on privacy
- On-device voice processing, where users’ voices are directly analyzed on the device they talk to, without anything going to the cloud.
The blockchain here is necessary as a way to incentivize clerks to do the job well, as well as to pay users for contributing their data. Without the token, there would be no reason for clerks and users to participate basically. And the fact that we use encryption throughout the process on top of decentralization means that we can guarantee privacy by design, while still training a neural network as if all the data was centralized an unencrypted!
BLOCK TRIBUNE: Since AI has a learning curve, won’t it still be susceptible to some errors, even in a decentralized system?
RAND HINDI: Decentralized AI has all the same issues as centralized AI, minus the privacy one. The objective here is to solve privacy. Everything else can be solved similarly to centralized system, with better models, better data, etc..
BLOCK TRIBUNE: Tell me about the token curated app store. What do you see being available?
RAND HINDI: Snips AIR device that will be available in the marketing by end 2019 will come packed with useful apps: from asking about the weather to finding recipes to cook, controlling your home or playing music, many of the most common use cases will be built-in.
In addition, Snips AIR will go together with voice app store, where users can find new voice apps to install and developers can create and publish new apps.
Centralized app stores have a number of major issues: users have no privacy, developers pay enormous fees, developers have limited freedom on what they can build, developers get kicked out arbitrarily,developers are being intermediated, rankings are opaque and gamed, reviews are often unreliable.
In designing the Snips voice app store, we followed two key principles: having no central authority and distributing most of the value created to the community. This is why using a token makes so much sense: it is a way to create a self-governed community that is incentivized to curate the app store and make it better.
The token-curated app store will have 3 types of actors involved: developers and companies publishing apps on the store (publishers), users of apps and users responsible for allowing apps on the store, flagging bad reviews and malware (curators).
Each user has a different incentive: the developers want to create and monetize apps, users want to do useful things with their assistants without impacting their privacy, and curators want to make sure the quality of the content remains high.
One of the main tools for creating such decentralized communities is called “staking”. In a nutshell, a participant in the economy wanting to access a feature must lock tokens, which can then be used as collateral in case they misbehave. Staking can be seen as a mix of behavioral economics and gamification, and simultaneously promoting good behavior in self-organized communities while also creating value for the token by acting as a velocity sink.
BLOCK TRIBUNE: Are you doing an ICO? If you are, provide the when/where details, including the hold-backs.
RAND HINDI: We have opened our token sale to the community as of late August.
BLOCK TRIBUNE: Will there come a day when virtually every electronic device has some form of AI embedded? How far out is that?
RAND HINDI: This is in fact our mission: to put Snips in every device on the planet, so that using them is as effortless as possible. Eventually, we won’t even think about it, and technology will disappear.
BLOCK TRIBUNE: I have to ask you the “Hal” question – some observers feel AI is a threat. Why are they wrong?
RAND HINDI: Artificial Intelligence will at best solve logical problems. It will never be capable of emotional intelligence, and, as such, will be complementary to humans, not replacing them! You can watch my latest TEDx talk on that topic as well here.