Intel’s Great Wall

Happy Monday and welcome to Patent Drop! 

Today, we’re checking out tech from Intel to put a 20-foot fence around AI models, Adobe’s tech to whip neural networks into shape, and a filing from PayPal for plans to prevent breaches using blockchain. 

Let’s take a look. 

#1. Intel’s data defender 

Machine learning models involve lots of time, energy, effort and especially data to create. So Intel understands that many might want to keep their hard work to themselves. 

The company filed a patent application for tech “secure model generation and testing” to protect the privacy of machine learning models and the data used to train them. Intel’s system allows a user to create, train and store a machine learning model in a “hardware protected space,” or a physical system that “protects the ML models within various hardware architectures to reduce the probability of attacks.” It modifies parts of the personally identifiable information used to train the model to reduce the probability that it’ll be identified as a bad actor. 

To get access and release the machine learning model, this system performs “distributed and multi-party analysis” to verify that the person attempting to gain access cannot see any of the data that the model was trained on. 

Essentially, Intel has created a vault to securely store an AI model that can be accessed without simultaneously releasing the data that’s been fed to it during training. As a result, an AI model can be “freely distributed” without concern of personal or identifiable data getting in the wrong hands. 

“ML models may be trained on sensitive and personal real-world data such as personally identifiable information,” Intel said in its filing. “Large scale collection of sensitive and personal data … entails risks. The consequences of mishandling sensitive data may be devastating from a privacy, business and economic perspective.” 

Keeping AI training data safe is crucial, given that models can often be trained on things like names, addresses, phone numbers or even drivers license and social security numbers, said Kevin Gordon, VP of AI technologies at NexOptic. 

Gordon noted that if a bad actor gets their hands on an AI model created by another party, it’s possible for them to reverse engineer it and gain access to the data that the model was trained on. But implementing a physical barrier that keeps said data locked inside can mitigate this.  

“These ML models are super data hungry, and usually they are hungry for our private data,” said Gordon. “If an attacker gets a hold of a black box AI model, they can regenerate some of those inputs that went into the training process. So if you’re putting in sensitive information, that might leak through poking and prodding of it.

We’ve already seen that patents protecting machine learning models themselves are pertinent to developers (For more on that, check out the Booz Allen patent application to watermark AI models that we broke down in Patent Drop last week). An AI model is only as good as the data it’s training on, so keeping that data safe is important to safeguard both personal privacy and companies’ IP (which, if we’re being honest, is likely what they care about more). 

While Intel lags behind in AI chip sales compared to the likes of NVIDIA, hardware to keep machine learning models secure may be part of the company’s plan to claw back some market share. If patented, Intel could become not just a provider, but the provider of this type of AI data protection in an exponentially growing market.  

#2. Adobe schools its neural networks

Neural networks are like kids: They take tons (and tons and tons) of resources to grow. Adobe wants to get you from Pre-K to college in a few short steps.

The company filed a patent application for “resource-aware training” for neural networks. To break it down: This tech works by identifying which neurons in a neural network are “dead” or “survived,” with dead neurons being defined as those “which have little or no impact on the network’s ability to reliably accomplish a task.” Those dead neurons are converted to “reborn” neurons by essentially weaving them in with the survived neurons. 

If you’re wondering how this saves on resources, the training system makes every neuron in a neural network as efficient as it can be. That means weeding out dead weight and retraining to be better, ergo, not wasting computing power on training parts of the neural network that won’t help it complete its task at the end of the day. 

If you’re still confused, think of it like this: If your plant is wilting on your desk inside, don’t waste time and water hoping it’ll spring back to life just because your other desk plants are thriving in the same environment. Just go plant it in a garden. 

The process of pruning a neural network for the parts that aren’t working has been around for decades, NexOptic’s Kevin Gordon told me. But Adobe is taking a waste not, want not approach by reintegrating and retraining the failed neurons to support the network’s function. 

“In the pruning step, you have a network with 100% capacity, but you start trimming off the neurons you don’t need … because they’re not contributing,” said Gordon. “Instead, they identify the kind of the subnetwork that they would prune, and train it specifically … to use the network to its fullest capacity. I can definitely see that being novel.” 

As Adobe noted, this process can also save on computational resources, which is an increasingly important factor given that the training of AI models is not so great for the environment. AI models consume massive amounts of energy, and training a single one can produce hundreds of thousands of pounds of CO2 emissions. Saving energy in any part of the process can ease the environmental strain, even if just by a bit. 

So why do neural networks matter so much to Adobe? This type of AI is a big part of completing “visual tasks such as image recognition and object detection,” the company said in its filing – And it is working hard to be a leader in AI as it relates to photo and video. The company has sought several patents for integrating AI into its suite, including a neural network-operated beauty editor and machine learning models for improved file rendering, and recently launched its own generative AI tool called Firefly, which the company announced Monday will be integrated into its video and audio editing software. Plus, the company has been “strong on the research side” of visual AI for the past decade, Gordon added. 

Adobe may be latching on to AI amid pressure to stay ahead of competitors like Canva and Figma … at least before the company bought Figma for $20 billion last year, Gordon said. “There’s huge threats to (Adobe) in the marketplace,” he said. “They have kind of the de facto standard on a lot of these creative tools… But now with these AI tools, it really takes them to the next level.” 

#3. PayPal’s (b)lock and key

PayPal wants to use blockchain for more than just crypto trading and/or NFT purchases. 

The company is seeking to patent a system that uses blockchain to verify and authenticate identity. First, a user uploads their credentials, such as username, password, address, phone number and credit card information, to be stored on a blockchain. When the user wants to access them in order to process transactions, the system generates a “user public key,” which can be used to access those credentials. 

Credentials are stored on what PayPal calls an “identity token.” This token, which comes with certain constraints (i.e., must be used to log in within a few minutes, or they can only be used for certain transactions) based on the situation, is then sent to the transaction server. 

Think of it as a more secure version of a digital keychain or password manager: This process adds additional layers of defense in storing personal information that could violate the user’s privacy if leaked. 

“With online identity theft increasing each year, theft of user credentials becomes a higher security concern,” PayPal said in its filing. “What is needed is a system that stores user credentials more securely than conventional systems and provides targeted credentials to authenticate the user.”

There are lots of ways that blockchain technology can be implemented in security, including “data immutability, decentralization, and strong cryptography,” Marius Grigoras, CEO of crypto startup launchpad BHero, told me. Blockchain itself fosters transparency with public transaction ledgers that can track fraud, and its decentralized architecture can minimize the risk of “single point of failure,” he added. 

Looking at PayPal’s patent application, blockchain can create a “tamper-proof and verifiable record of user credentials, which allows for seamless authentication without relying on intermediaries,” Grigoras said, which has the potential to “drastically” cut down data breaches. 

The use of blockchain in security this way has the potential to reach far beyond financial services companies, Grigoras noted. This tech can be applied to any industry that deals with confidential data, including supply chain management, healthcare, and social media. 

Still, fintechs like PayPal tend to have access to a wealth of important personal and private data. PayPal itself acknowledged this in its attempt to patent a “personal data wallet” in recent months with the capability to store everything from your credit card information to your birthday and social security number. 

But notably, PayPal’s track record with personal data security faced fire earlier this year amid a large-scale security breach that left more than 35,000 users’ credentials vulnerable. If the company wants to be a personal data stronghold, bolstering its defenses is vital. 

Have any comments, tips or suggestions? Drop us a line! Email at admin@patentdrop.xyz or shoot us a DM on Twitter @patentdrop. If you want to get Patent Drop in your inbox, click here to subscribe.