About the project

The inspiration for GuardianEdge came from the growing concern over privacy in home and commercial security. Most modern smart cameras rely on cloud-based AI, which not only introduces latency in critical threat detection (like fire or weapons) but also uploads sensitive video data to remote servers. We wanted to build a "blind" security system—one that is intelligent enough to detect threats locally but never needs to send a single pixel to the cloud.

Inspiration

Security should be private. We believe that AI can be powerful without being invasive. By moving intelligence to the edge, we eliminate the need for cloud streaming and protect user data from potential breaches.

What it does

GuardianEdge is a real-time security application that uses a PerforatedAI-optimized YOLOv8 model. It performs on-device object detection to identify specific threat classes—such as knives, firearms, or fire—while ignoring non-threatening activity.

By leveraging dendritic optimization, the system can run on less powerful edge hardware with higher efficiency than standard large-scale models. When a threat is detected, it triggers a local alert and log without compromising the user's data privacy.

How I built it

The project is built on the Ultralytics YOLOv8 framework. The core innovation is the integration of PerforatedAI's dendritic optimization.

1. Model Architecture: I used YOLOv8n (nano) as the baseline for high speed on edge devices.
2. PAI Integration: Using perforatedai, I initialized the YOLO backbone with dendritic structures.
3. Training: I implemented a custom training loop that uses Perforated Backpropagation (PB) to find the most "important" neurons and grow dendrites around them. This is represented by the formula: $$W_{new} = W_{old} + \Delta W_{PAI}$$ where $\Delta W_{PAI}$ is optimized via the Dendritic tracker.
4. Inference: A Python-based real-time wrapper using OpenCV handles the video stream and overlays the optimized detections.

Challenges I ran into

Integrating a custom optimization framework into the highly encapsulated ultralytics training pipeline was the biggest technical hurdle. I had to bypass the standard .train() method and write a manual loop that correctly managed the GPA.pai_tracker callbacks at the end of each epoch to ensure the dendrites actually "grew" based on the validation scores.

Accomplishments that I'm proud of

I am incredibly proud of achieving a stable dendritic growth pattern that showed measurable improvement in the PAI validation graphs. Seeing the "PB Scores" rise for specific convolutional layers proved that the optimization was surgically targeting the parts of the model that needed help to run better on edge hardware.

What I learned

I learned that "smaller is often better" when it comes to edge AI. By using dendritic optimization, you can make a nano-sized model behave with the precision of a much larger one. This is a game-changer for privacy-focused hardware.

What's next for it

The next step is to port the optimized models to specialized hardware like the NVIDIA Jetson or Coral TPU. I also plan to implement "Secure Enclave" storage where even the local logs are encrypted with a key only the user holds.

Built With

• perforatedai
• python
• ultralytics
• yolov8