I recently put together a TensorFlow visual recognition harness for some security research I’m doing in regards to my long standing Bird Hunting Season Project.
I wanted to be able to see what the Falcon/Sparrow/Flex and the Picard/Bravo Compute Box sees and recognizes without keeping the devices on all the time.
So I put together a harness that uses the same pipeline, models and flow straight from the device. Will be interesting to see if anything comes from it.
I thought I’d be able to release the harness as you see in this post just without the models but unfortunately I can’t.
So instead I released a TensorFlow Generic Harness that supports whatever you bring to it. Link
Its a simple toolkit that runs YOLO/SSD TensorFlow Lite models against live webcams, recorded footage or extracted session directories. While mirroring what you might see on devices with JSON reporting, session markers. I even added Docker and venv setup automated for MacOS and venv.
It is bring your own model assets though!
Was definitely a cool thing to spin up. We’ll see if anything comes from it or if it just ends up getting added to BirdShot (which I promise I will at least show off if not release one day!)
Part 1: Bird Hunting Season – Security Research on Flock Safety’s Anti-Crime Systems: HERE
Part 2: Plucked and Rooted – Device 1: Debug Shell on Flock Safety’s Raven Gunshot Detection System: HERE
Part 3: Grounded Flight – Device 2: Root Shell on Flock Safety’s Falcon/Sparrow Automated License Plate Reader: HERE
Part 4: Trap Shooter – Flock Safety Sniffer & Alarm: HERE
Part 5: Root from the Coop – Device 3: Root Shell on Flock Safety’s Bravo Compute Box: HERE
Part 6: Fly-By – Device 2: The Falcon/Sparrow – Gated Wireless RCE, Camera Feed, DoS, Information Disclosure and More: HERE
Part 7: Button Presses to Wireless RCE: Shell on Flock Safety’s License Plate Cameras Over Wi-Fi: HERE
Part 8: Formalizing my Flock Safety Research: HERE
Part 9: BirdEye (Tool to Test Flock Safety’s ML Visual Recognition Models): HERE
END TRANSMISSION