Smarter Cameras Need More Than Edge AI to Protect Privacy

Axis Communications, a professional security camera manufacturer with approximately 5,500 employees and about 90,000 system integration partners, is embedding AI analysis directly into its cameras via its ARTPEC system-on-chip so that privacy masking happens before video leaves the device, according to Mats Thulin, the company's director of AI and analytics solutions. Thulin told EE Times that cameras should not "collect and record more material or data than what's needed to fulfill the use case," with applications such as Axis Live Privacy Shield detecting people, faces, or license plates and replacing those pixels in-camera, while legitimate incident-evidence use cases at banks, schools, retailers, airports, and transport operators retain an encrypted, unmasked stream in protected storage. Axis also cryptographically signs footage and, in some products such as acoustic sensors that detect screams, shouts, or breaking glass, prevents conversation streaming at the device level.

A camera monitoring a distribution center may need to detect a blocked emergency exit, warn when a worker enters a hazardous area, or follow packages through a production flow. None of those tasks require knowing the worker’s identity. Yet that same camera may need to preserve recognizable evidence if a theft, accident, or intrusion occurs. The challenge for intelligent video systems is therefore not simply to see less; it’s to determine when identity is necessary, suppress it when it’s not, and tightly control the footage when it must remain. “The idea is not to collect and record more material or data than what’s needed to fulfill the use case,” Mats Thulin, director of AI and analytics solutions at Axis Communications, told EE Times. Intelligence moves into the camera Axis is a professional security camera manufacturer, with approximately 5,500 employees and a network of about 90,000 system integration partners. Its cameras increasingly act as real-time sensors rather than passive recording devices. View All https://www.eetimes.com/category/sponsored-content/ Thulin described systems that detect people entering hazardous areas, identify missing hard hats, warn of pedestrians near industrial vehicles, monitor blocked exits, and provide early warnings of smoke or overheating equipment. Cameras can also count occupants, track packages, and support manufacturing quality assurance. “In all of those examples, you blur out people to protect privacy because you’re not interested in recording video of people,” Thulin said. Axis cameras perform AI analysis locally using the company’s ARTPEC system-on-chip SoC . According to Thulin, all cameras in the current portfolio are AI-enabled, with deep-learning processing handled directly in the ARTPEC SoC rather than requiring separate dedicated hardware. Applications such as Axis Live Privacy Shield analyze the stream inside the camera; detect people, faces, or license plates; and replace the relevant pixels before routine video leaves the device. Fixed areas, such as windows overlooking private property, can also be blocked. For an occupancy application, the camera might send only a count to a dashboard. In a distribution center, it can follow parcels while masking employees. Processing locally also reduces the need to transmit continuous video to the cloud. “Instead of feeding the full video to a cloud environment where it’s stored somewhere, and you lose track of privacy, the camera can perform the counting and send only the required result,” Thulin said. But not every camera is used only for counting or process monitoring. Banks, schools, retailers, airports, and transport operators may have legitimate reasons to retain identifiable evidence following an incident. Thulin described a bank installation in which routine users receive masked video, while the camera also sends an encrypted, unmasked stream to protected storage. Operator and administrator roles can separate routine monitoring from the ability to change privacy settings or retrieve identifiable footage. “If the use case is to identify a break-in, of course you want to record the person doing an intrusion,” Thulin said. “In those cases, it’s about how you maintain the security of that data.” Axis also cryptographically signs footage so an organization can demonstrate that it came from a particular camera and was not subsequently altered. Long-term software support, patching, encryption, and access controls are therefore part of the privacy architecture. In some cases, Axis restricts the device itself. Thulin pointed to acoustic sensors that can detect screams, shouts, or breaking glass but cannot be configured to stream conversations. Privacy is stronger when unnecessary surveillance is technically unavailable rather than merely prohibited by policy. Privacy risks continue after recording But even carefully controlled footage can become intrusive when it’s extracted and shared. Pimloc addresses this later stage of the video lifecycle through its Secure Redact platform. The software uses AI to detect and track faces, heads, bodies, license plates, scene text, and other sensitive details across recorded video. An authorized user can choose which individuals should remain visible, anonymize everyone else, review the results, and generate a separate file for release. Simon Randall, CEO of Pimloc, gave the example of a school sharing footage of an incident with parents. The parents may need to see what happened to their child but should not receive identifiable video of every other child present. “The anonymization needs to be irreversible,” Randall told EE Times. “We effectively create a whole new version of the file with that personal information removed. So it’s theoretically impossible to get back to what was there before.” Users may blur or pixelate faces or replace an entire region with black pixels. Full bodies can be removed when a person’s height, clothing, or surroundings could reveal identity. Secure Redact can also remove names, addresses, financial information, and individual speakers from audio, while stripping metadata that could disclose time or location. The original is not necessarily destroyed. It may remain in protected storage for evidentiary purposes, while only the redacted version is distributed. “If you anonymized it at the device, it’s better—it’s more private—because you never capture the stuff to start with,” Randall said. “The question is: Under what circumstances is that useful?” Pimloc automates tasks that can otherwise be prohibitively slow. Randall cited a customer that required three-and-a-half weeks to manually redact a particularly challenging 20-minute body-camera video. He said Pimloc reduced the review stage to about 15 minutes. But automation does not eliminate human verification. Real security footage may be grainy, crowded, poorly lit, or captured by a rapidly moving body-worn camera. Users must confirm that every relevant identifier has been found and that the correct people remain visible. The platform also preserves a record of what was changed. Pimloc hashes the incoming file, logs the redaction process, and hashes the finished version. That chain of custody complements Axis’s camera-level signing. “The other thing that’s just as important now is being able to prove authenticity and provenance of content,” Randall said. “There’s privacy in not disclosing private information, but the other thing is making sure that what you do disclose is true and accurate.” Privacy across the lifecycle Strong privacy can make video more usable rather than less usable. Properly anonymized footage can be shared with parents, citizens, lawyers, operational teams, researchers, or the public without unnecessarily identifying bystanders. “I would argue that it’s a counterintuitive point, but by having really good privacy, it means you can actually make a lot more use of the data,” Randall said. Neither company, however, treats edge AI or masking as sufficiently on its own. Cameras can be reconfigured, administrators can misuse access, vulnerabilities can expose streams, and a system installed for safety can gradually expand into employee monitoring or customer profiling. Cybersecurity, access control, retention limits, auditing, integrator practices, and regulation determine whether the original privacy design survives. “A lot of people put privacy and security at opposite ends of a continuum,” Randall said. “We need to do both together.” Security cameras can therefore become more intelligent without necessarily becoming more intrusive. But privacy must extend across the full lifecycle: what the camera captures, what it analyzes, what it transmits, what it stores, who can retrieve it, and what remains visible when it’s shared. The decisive question is not simply whether AI runs at the edge; it’s whether the system can justify when identity matters, technically suppress it when it does not, and protect the original whenever it must remain. Read also: The Security AI That Learns the Language of Movement https://www.eetimes.com/the-security-ai-that-learns-the-language-of-movement/ Ambient IoT: From Battery-Free Promise to Mass-Market Reality https://www.eetimes.com/ambient-iot-from-battery-free-promise-to-mass-market-reality/ India Hardens Its Cyber Defenses https://www.eetimes.com/india-hardens-its-cyber-defenses/ LEDs Push Wireless Power Further https://www.eetimes.com/leds-push-wireless-power-further/