Deep Defender

Real-time defense against deep fakes and malicious editing.

This is an experimental proof of concept app. No one should rely on this yet. I think the core idea is sound, but I would not be surprised if this particular implementation is possible to circumvent, since the audio fingerprinting algorithm I'm using is pretty weak and the app itself is hacked together. I would welcome any input from security and ML experts to make the app more robust.

How it works

To use the app, just run it on your phone while you're being filmed, and make sure the screen is visible to the camera. You could prop it up on the desk in front of you, or maybe put it in your shirt pocket.

As you speak, the app listens to what you're saying, and every second or so it generates a fingerprint of what it hears. Then it cryptographically signs the fingerprint with your private key and displays it as a QR code. This is called a Signed Audio Fingerprint (SAF code).

Anyone can analyze the video to generate the audio fingerprint, and verify that it matches the fingerprint in the SAF code. They can also verify the signature using your public key.

SAF code format

[        signable data       ] [signature]
[metadata] [audio fingerprint] [signature]

[metadata] = [magic string] [version] [algorithm] [time]
magic string = "SAF"
version = Varint
algorithm = Varint
timestamp = Uint64 (big endian), ms since unix epoch

At the moment the version and algorithm IDs are hard-coded to 1. The signature algorithm is ECDSA-sha256.

The fingerprint currently includes a robust audio hash, and the RMS volume of the chunk, totalling 100 bytes.

For more details, check out the blog post.

Building

Use the flutter tool to build and run in the usual way:

flutter pub get
flutter run

And run the tests like this:

flutter test

Running the verifier

The verifier has two stages. The first stage extracts all the SAF codes from the video, and also extracts the audio track. The second stage verifies all the SAF codes by comparing them to the audio.

The reason these stage are separate is that I couldn't find a good QR code reading library for Dart. So the first stage is a python script. This script depends on the pip packages ffmpeg-python, numpy, and zbar-py, so install those first. Then run:

python3 bin/read_saf_codes.py your_video_file.mp4

This will output two files: your_video_file.mp4.wav and your_video_file.mp4.safcodes.txt. Pass these to the second stage, along with the target's public key:

dart run bin/run_verifier.dart your_video_file.mp4.wav \
    your_video_file.mp4.safcodes.txt public_key.json

The public key file is encoded in JWK (JSON Web Key) format, which is the same format exported by the "Share public key" button in the app.

run_verifier.dart will print the validity of each SAF code, but it's currently a very bare bones script. It needs to be updated to print a richer report.