Pythia Camera
Diagram: User Flow
Diagram: AWS
Diagram: Chained RSA Signature

Pythia Camera

Check out the github.

Pythia Diagram

Inspiration

Original Idea:

Deepfakes and more standard edits are a difficult threat to detect. Rather than reactively analyzing footage to attempt to find the marks of digital editing, we sign footage on the camera itself to allow the detection of edited footage.

Final Idea:

Using the same technology, but with a more limited threat model allowing for a narrower scope, we can create the world's most secure and intelligent home security camera.

What it does

Pythia combines robust cryptography with AI video processing to bring you a unique home security camera. The system notifies you in near-real-time of potential incidents and lets you verify by viewing the video. Videos are signed by the camera and the server to prove their authenticity in courts and other legal matters. Improvements of the same technology have potential uses in social media, broadcasting, political advertising, and police body cameras.

How we built it

Records video and audio on a camera connected to a basic WIFI-enabled board, in our case a Raspberry Pi 4

At regular intervals:

Combines video and audio into .mp4 file
Signs combined file
Sends file and metadata to AWS

Signing

On AWS:

Verifies signature and adds server signature
Uses Rekognition to detect violence or other suspicious behavior
Uses Rekognition to detect the presence of people
If there are people with detectable faces, uses Rekognition to
Uses SMS to notify the property owner about the suspicious activity and links a video clip

Challenges we ran into

None.

Just Kidding:

Hardware

Raspberry Pi

All software runs on Raspberry Pi
Wifi Issues
Compatibility issues
Finding a Screwdriver

Hardware lab didn't have the type of sensors we were hoping for so no heat map :(.

Software

Continuous batched recording
Creating complete .mp4 files
Processing while recording

Web Services

Asynchronous Architecture has lots of race conditions

Accomplishments that we're proud of

Complex AWS deployment
Chained RSA Signature
Proper video encoding and processing, combining separate frame and audio streams into a single .mp4

What we learned

Bogdan

Gained experience designing and implementing a complex, asynchronous AWS Architecture
Practiced with several different Rekognition functions to generate useful results

Philip

Video and audio encoding is complicated but fortunately we have great command-line tools like ffmpeg
Watchdog is a Python library for watching folders for a variety fo events and changes. I'm excited to use it for future automation projects.
Raspberry Pi never works right the first time

What's next for Pythia Camera

A lot of work is required to fully realize our vision for Pythia Camera as a whole solution that resists a wide variety of much stronger threat models including state actors. Here are a few areas of interest:

Black-box resistance:

A camera pointed at a screen will record and verify the video from the screen
Solution: Capture IR footage to create a heat map of the video and compare the heat map against rekognition's object analysis (people should be hot, objects should be cold, etc.
Solution: Use a laser dot projector like the iPhone's faceID sensor to measure distance and compare to machine learning models using Rekognition

Flexible Cryptography:

Upgrade Chained RSA Signature to Chained RSA Additive Map Signature to allow for combining videos
Allow for basic edits like cuts and filters while recording a signed record of changes