Synthetic faces, real data: Syntonym and Zenseact tackle autonomous driving's privacy gap

For engineers teaching autonomous vehicles to navigate city streets, the most valuable resource is data. A vehicle needs thousands of hours of video footage to recognise a pedestrian's gait or a cyclist's glance. But collecting that footage creates a legal problem: much of it contains people's faces and license plates, which privacy regulations requires companies to protect. 

The standard fix is to blur them. For a machine-learning model, though, a blurred face removes exactly the information that matters. The direction of a gaze, a subtle shift in expression, the moment a person decides to step into traffic: blurring strips all of that away. 

That tension between privacy compliance and data quality is the focus of a new project between Zenseact, one of MobilityXlab’s partner companies, and Syntonym, a startup from the program’s Batch 9. Together, they are working to bring face anonymisation to one of the industry's largest open datasets.

Replacing faces, not erasing them 

Syntonym's technology uses generative AI to replace real human faces with synthetic ones. The digital substitutes are not identifiable, but they preserve the physical characteristics that matter for autonomous driving: the exact angle of the neck, the dilation of pupils, micro-expressions around the mouth. To a privacy regulator, the person is gone. To the vehicle's perception system, the person's intent is still legible. 

Batuhan Özcan, Syntonym's founder, calls this approach creative construction. Rather than treating privacy compliance and data quality as opposing forces, the idea is to build the safeguard directly into the data from the start. 

"We are very pro-responsible AI," he said. "We are not destroying the data to protect the person. We are building a safeguard directly into the pixels." 

The immediate aim is to enhance the Zenseact Open Dataset, a large multi-modal autonomous driving dataset built by Zenseact researchers and made available to the global research community. With synthetic anonymisation, the dataset will go beyond current industry standards for data utility. 

"Zenseact's precautions will be way more elegant than other alternatives," Batuhan said.

Timing is everything 

Syntonym joined MobilityXlab in 2023, as part of Batch 9. At the time, the technology was ready, but Zenseact had a different timing. In the world of deep-tech startups, this gap is where some companies can fail. As Batuhan explains, “in large organizations, the need has to become inevitable before the contract is signed”. 

Rather than stepping back, Syntonym stayed in contact with Zenseact over three years, showing up at events and keeping the relationship active. This year, Zenseact's internal roadmap for the Open Dataset finally aligned with what Syntonym had been building. 

When asked what he would tell other founders on the same path, Batuhan was direct: acceptance into MobilityXlab is a strong signal of validation, but it is only the beginning. He encourages startups to treat these relationships as long-term, to keep building their customer base outside the program, and to stay present. "Never give up," he said, "because there might be another time that stars align."