Do you think that AI systems can make security systems smarter without real data? This is what companies are experimenting with extensively. They are finding ways to make security systems better with the help of synthetic data.

Verkada is known for making security cameras and other safety devices. It runs a huge network of connected devices, almost 2.4 million. This includes cameras, alarms, sensors, intercoms, access control systems, and other related security devices. Their devices are used by 30,000 companies. And above all, their network is spread across 170 countries. Now Verkada has announced a deal with NVIDIA. As a part of this deal, NVIDIA will be investing money in Verkada, though the exact amount has not been declared yet. Besides that, both companies have also agreed to work together in the field of technology.

What’s Happening

As a part of this deal, NVIDIA will be providing two of its technologies to Verkada: Cosmos foundation models and Physical AI Data Factory. As far as the second feature from NVIDIA is concerned, it basically generates fake videos. You must be thinking, what is the purpose of generating these fake videos? Well, these are super helpful for training an AI model. If the original videos are missing, AI models can be trained by creating these synthetic or fake videos. Sounds interesting? Yeah, this is what NVIDIA is providing to Verkada as part of the partnership.

Verkada is using these features to improve its video search feature. How does video search work? This works when you have hours of recordings and you need to find a particular person or object in the footage. Will you be watching hours of footage? Of course not. This is what Verkada is working on. They said that after this new approach, they have seen a 68% improvement in the accuracy of spatial-temporal search within their video search feature, and that is incredible.

Verkada’s co-founder and CEO, Filip Kaliszan, said the company has been working on physical AI for almost ten years, even before people commonly used that term. He also stated that this partnership will help them keep students and factory workers safe while also stopping retail theft. 

Why NVIDIA is Doing This

This isn’t NVIDIA’s first move into this space. For roughly two years, the company has been building relationships with businesses in robotics, self-driving vehicles, and factory automation, using its Cosmos models and related developer tools. It has also backed several startups that build their products on top of NVIDIA’s technology.

Verkada fits well into this strategy because it already has years of camera footage and data. Combining NVIDIA’s pre-built AI models, synthetic data tools, and fast processing systems helps solve two common problems in video-based AI: not having enough real examples of rare situations, and not having enough computing power to process everything quickly in real time.

What This Means for People Building Similar Systems

For engineers and companies working on similar video AI systems, this deal points to a few useful lessons.

First, synthetic data can genuinely help systems get better at spotting rare or unusual situations. But this only works well if teams also keep testing the system against real, untouched footage to check that it’s actually improving and not just performing well on made-up data.

Second, using large foundation models to understand video content usually increases the amount of work needed from vector databases, which store and search through data. What does that mean? This means teams should plan ahead for slower search speeds and higher costs as they scale up.

Third, running AI across millions of devices brings its own challenges, like deciding which processing should happen directly on the device versus in the cloud, how to safely roll out software updates, and how to keep data secure as it moves across the network.

It’s also worth noting that the 68% accuracy improvement is based on internal testing by the two companies. Results like this can vary a lot depending on the type of data, the kind of search being done, and how the testing was set up. So far, neither company has released a public, repeatable benchmark that others can verify.

What to Watch Next

Besides all the positives of this partnership, there are a few questions that remain open. Will NVIDIA release case studies or benchmarks that researchers can implement if they want to? Will Verkada show that its synthetic data will perform better over time? Will this data be used as effectively as live footage? And above all, there are privacy concerns too. Will the company be able to protect the privacy of users? Will schools’ and factories’ data be safe? Will there be misuse of AI? Users may be more concerned about it because of the security problems Verkada faced earlier in 2021, when its camera feeds were hacked.