Extend verification to attention-based and multimodal semantic communication systems

Investigate the application of the neural-network verification and certifiable-robustness framework to attention-based and rate-adaptive deep joint source-channel coding successors, including WITT and SwinJSCC, and to communication modalities beyond images, using abstract domains capable of representing vision-transformer attention layers.

Background

The paper develops a bound-propagation framework for verifying deep joint source-channel coding decoders with PReLU activations, substituted transposed convolutions, and AWGN or Rayleigh-fading channel models. Its treatment is designed around convolutional architectures and linear relaxations for ReLU-family activations.

The authors identify extension to newer attention-based and rate-adaptive systems, specifically WITT and SwinJSCC, as unresolved because their vision-transformer attention layers require abstract domains beyond those used in the present framework. They also explicitly include modalities beyond image transmission within the open direction, and point to abstract domains for large GPT models as a possible foundation.

References

Applying this approach to attention-based and rate-adaptive successors, such as WITT and SwinJSCC, as well as to modalities beyond images, remains an open research direction. These successors are based on vision transformers (ViTs), whose attention layers require abstract domains beyond the \relu{} and \prelu{} relaxations used in this work. Recent abstract domains developed for verifying large generative pre-trained transformer (GPT) models provide a promising foundation for this extension.

— Neural Network Verification for Deep Joint Source-Channel Coding  (2610.11994 - Le et al., 8 Oct 2026) in Section Conclusion