Privacy Masking in Security Cameras Uses AI to Protect Innocent Bystanders

Privacy masking in security cameras has evolved from crude static black boxes drawn over portions of the image to sophisticated AI-driven systems that automatically detect and blur faces, licence plates, and neighbouring properties in real time. This technology resolves one of the most persistent tensions in modern security — the need for comprehensive surveillance coverage versus the legal and ethical obligation to protect the privacy of innocent bystanders, neighbours, and passersby who are incidentally captured by cameras installed for legitimate security purposes.

For New Zealand property owners and security operators, AI-powered privacy masking represents a technological solution to what has been a regulatory and social challenge. It allows cameras to watch everything they need to watch while automatically ensuring that people who are not the subject of the security monitoring have their privacy preserved — all without any manual intervention or ongoing management.

How AI Privacy Masking Works

Traditional privacy masking used fixed zones — rectangles or polygons drawn on the camera’s configuration interface that permanently blacked out specific areas of the image. While useful for blocking views of neighbouring windows or public spaces, static masks are rigid. They cannot adapt to moving subjects, cannot distinguish between security-relevant individuals and innocent bystanders, and cannot be applied selectively based on context.

AI-powered dynamic privacy masking is fundamentally different. It uses real-time object detection and classification to identify specific elements in the video stream and apply masking to them automatically.

The process works as follows:

• Frame analysis: Every video frame passes through an AI model that detects and classifies objects — faces, licence plates, people, vehicles
• Selective masking: Based on configured rules, the system applies blur, pixelation, or silhouette masking to selected object types while leaving the rest of the image in full resolution
• Real-time processing: The masking is applied at the frame level before the video is stored or transmitted, ensuring that unmasked footage never exists in the recorded stream
• Dual-stream recording: Advanced systems record two streams — a privacy-masked stream for general access and a full-resolution stream stored with enhanced encryption for authorised forensic access only

The AI processes video at frame rate — typically 25 to 30 frames per second — ensuring that masking is seamless with no gaps or flicker. The masked elements are tracked across frames, so a face that enters the camera’s field of view is masked immediately and remains masked throughout its transit, regardless of how the person moves or turns.

Types of AI Privacy Masking

Different masking techniques serve different purposes, and modern systems often support multiple methods that can be applied based on the use case.

Face Blurring

The most common application of dynamic privacy masking is automatic face blurring. The AI detects faces in the video stream and applies a blur or pixelation effect that renders the face unrecognisable while preserving the rest of the person’s body and the surrounding scene. This approach is widely used in public-facing camera deployments where security monitoring is necessary but individual identification is not the primary purpose.

Licence Plate Masking

Similar to face blurring, licence plate masking automatically detects and obscures vehicle registration numbers. This is particularly relevant for cameras that capture public roadways or car parks where recording the plate numbers of uninvolved vehicles creates unnecessary personal data collection.

Full Body Silhouette

Some privacy masking systems replace detected persons with coloured silhouettes — preserving the ability to see that a person is present and track their movement through the scene while completely obscuring their appearance, clothing, and physical characteristics. This approach provides maximum privacy protection while maintaining useful security monitoring capability.

Neighbouring Property Masking

AI can be trained to identify and mask areas of the image that belong to neighbouring properties. Unlike static masks that block a fixed region, AI-based property masking can adapt to camera movement (on PTZ cameras) and distinguish between the monitored property and adjacent ones based on learned spatial boundaries.

Legal and Regulatory Context in New Zealand

New Zealand’s Privacy Act 2020 establishes that personal information — including identifiable images — should be collected only when necessary and stored only as long as required. The Office of the Privacy Commissioner has received numerous complaints about security cameras that capture more than they need to, particularly in situations where cameras overlook public spaces, neighbouring properties, or areas where people have a reasonable expectation of privacy.

AI privacy masking directly addresses the Privacy Act’s proportionality requirements. By automatically minimising the capture of identifiable information about uninvolved individuals, the technology demonstrates that the camera operator has taken reasonable technical measures to limit privacy intrusion — a key consideration in any privacy complaint or investigation.

The technology also supports compliance with the Act’s purpose limitation principle. A security camera installed to monitor a commercial entrance does not need to record the faces of every pedestrian walking past on the public footpath. AI masking ensures it does not, while still providing full security coverage of the entrance itself.

Dual-Stream Architecture for Forensic Access

A sophisticated implementation of AI privacy masking uses a dual-stream approach that balances privacy protection with forensic capability. The primary stream — used for live monitoring, routine review, and general access — has full privacy masking applied. All faces, plates, and other configured elements are masked before this stream is viewed or stored.

The secondary stream retains full-resolution, unmasked footage but is stored with enhanced security controls:

• Stronger encryption: The forensic stream uses a separate encryption key, accessible only to designated authorised personnel
• Access logging: Every access to the forensic stream is logged with the accessor’s identity, timestamp, and stated purpose
• Approval workflow: Some systems require multi-person authorisation to access the forensic stream, preventing any single individual from viewing unmasked footage
• Shorter retention: The forensic stream may have a shorter retention period than the masked stream, automatically deleting after a defined period unless flagged for preservation
• Legal hold capability: When footage is needed for an investigation, specific time periods can be flagged for preservation, preventing automatic deletion

This architecture ensures that if a security incident occurs and identification is genuinely needed, the full-resolution footage exists and can be accessed through proper authorisation channels. For all other purposes, only the privacy-masked version is available, ensuring that routine security monitoring does not create an unnecessary repository of identifiable personal data.

Implementation Considerations

Deploying AI privacy masking effectively requires attention to several practical factors.

Processing requirements for real-time AI masking are substantial. Each camera stream must be processed through the AI model at full frame rate, which requires either on-camera AI processing (edge-based) or a sufficiently powerful server for centralised processing. Edge-based masking is preferred for privacy because unmasked footage never leaves the camera, but it requires cameras with built-in AI capabilities.

Accuracy in challenging conditions must be verified. AI face detection can struggle with extreme angles, partial visibility, very small faces at distance, and unusual lighting conditions. Testing the system under real-world conditions at the deployment site is essential to confirm that masking is applied consistently.

Configuration of masking policies should reflect the specific deployment context. A camera monitoring a private car park may need licence plate masking but not face masking for employees. A camera overlooking a public footpath should mask all faces. A camera monitoring a retail entrance may mask faces during normal operation but disable masking in the forensic stream for theft investigation.

The Future of Privacy-Preserving Surveillance

AI privacy masking represents a fundamental shift in how we think about security surveillance. Rather than accepting privacy intrusion as an unavoidable cost of security, the technology demonstrates that comprehensive monitoring and meaningful privacy protection can coexist.

The most advanced security camera is not the one that captures everything — it is the one that captures what matters while intelligently protecting everything else. AI privacy masking makes this distinction automatically, continuously, and reliably.

For New Zealand’s security industry, AI privacy masking is transitioning from a premium feature to a standard expectation. As public awareness of surveillance privacy issues grows and regulatory enforcement becomes more active, the ability to demonstrate that cameras are technically configured to minimise privacy intrusion will become as important as the cameras’ ability to detect threats. The technology exists today to achieve both — and property owners who deploy it are positioned on the right side of both the law and public expectations.