Variational Quantum Neural Networks Restore 0.95 Accuracy in Noisy Semantic Communication on MNIST

The framework compresses classical data into semantic representations, encodes them via a variational quantum transmitter, sends them over a quantum channel, and decodes at a trainable quantum receiver. Experiments used the two-node case with ideal, bit-flip, depolarizing, and amplitude-damping channels. A baseline model trained on a perfect channel reached 0.9556 accuracy and 0.9551 F1-score before noise was introduced. Receiver-side retraining at fixed depolarizing levels substantially restored performance under moderate and high noise. Task-relevant semantic features proved recoverable even when the physical density matrix remained distorted, separating semantic recovery from full quantum state tomography. This result aligns with NISQ-era efforts to embed machine learning directly into quantum links rather than relying on classical post-processing. It suggests semantic communication could tolerate realistic channel errors if the receiver is jointly optimized, an approach already explored in classical 6G semantic-communication prototypes but rarely tested on actual quantum hardware. Next steps include scaling to multi-node topologies and real QPUs; without larger datasets and hardware validation, claims of near-term everyday deployment remain speculative.