Glow: Better Reversible Generative Models

I think he's suggesting that you play the video for the camera, and then let the camera sign what it saw.

> If GPS satellites were to cryptographically sign their signals, would that be sufficient to prevent spoofing of GPS signals (assuming the receiver of the signals treated the signing appropriately)?

Nope, classic replay attack case (just record the GPS signal and replay to the device at the desired location). You'd need a true time signal within the device, e.g. and atomic clock to make it work (so you'd authenticate the signed time against true time).

---

There is another way, however. If we assume the hardware is tamper-proof (unless drastically different methods are needed), then with strict timings we can device a challenge-response system that's immune to replay attacks due to relativity: simply transmit a signal A, have a known 3rd party (e.g. US government servers in cellphone towers) sign your signal Sig(A) and retransmit, and check that the delay matches the propagation delay you'd expect from the cellphone tower distance, plus the fixed (and immutable since it would be gov-controlled) processing delay. Your tamper-proof crypto-camera would record its location and whether it trusts the location. Using cellular signals is also better because GPS doesn't work indoors and is sensitive to interference.

Since we're adding a cellular connection to our device, it would also be a good idea to log its position on the state-controlled servers (again can be done with cryptographic safety assuming non-tampered device), along with some kind of intrusion detection system. As soon as it'd detect an attempt at tampering, it would relay this attempt to the servers, storing the intrusion and invalidating the authenticity subsequent recordings; a self-destruction attempt of the key would probably be also wise.

---

And now that I think about it, you'd probably want to put several keys/auths in the device, from different organizations -- not only governments. That way if the government authentication is positive but NGO's mismatch, you can suspect a government-backed forgery attempt (analogously vice-versa).

yes, thank you, I understand the motivation, the reason I never understood before was quite simple, the description "1x1 convolution" (correctly) described that they were not doing convolutions across pixels.

It is good to point out what they aren't doing, they should also point out what they are doing.

Humor me by considering:

"Today I did not go to the zoo, I did not go to school, ..., and I did not go by foot, nor by vehicle without weels, nor by a vehicle with only one wheel, nor by a vehicle that had 3 or more wheels, and the vehicle was not motorised, and I did not need to stand up on the vehicle"

While true, and possibly important to point out what I didn't do, it's generally more helpful to describe what I am doing, like "I rode my bicycle to the supermarket"

There is no mechanism in this paper (or in a standard VAE, or GAN) to encourage that a single human-understandable semantic quantity should be captured in a single dimension of the latent code. So, in general, it won't happen. "Blondeness" is spread out over all dimensions of the latent code. Therefore the method in this paper (summarized by matheist) is totally reasonable. There are other autoencoder-type generative models that try to concentrate each attribute in one dimension, usually by using the class labels as an additional input. But that is not the focus of this paper.