A verifier bounds image errors across wireless channels

DeepJSCC learns to send image information through a wireless channel and reconstruct it at the receiver. Training optimizes average quality on sampled channel conditions, while spot tests can catch particular failures but cannot show how bad reconstruction could get anywhere in a specified disturbance range. Can that worst-case error be bounded?

The paper builds a verifier that bounds a DeepJSCC decoder’s image error for every channel input inside a specified noise region.

The authors train three decoder sizes on CIFAR-10, a ten-class image collection, then use bound propagation to bound average absolute pixel error across every input in a specified channel region.

For the large decoder under AWGN, its certified upper bound stayed above 1 at every tested signal-to-noise ratio after ordinary training; further training with a sensitivity penalty lowered it to 0.08 at 30 dB. This shows certification became possible in the tested conditions.

The guarantee applies only to inputs inside the specified channel region and bounds error for the image instance being checked.

The radio test covered ten fixed images, one per CIFAR-10 class, at the measured channel conditions; it does not establish behavior over the full image test set or outside that channel envelope.