AI-based receivers demonstrate improved performance in high-noise environments and can reduce communication overhead compared to conventional receivers. However, their performance is significantly dependent on the representativeness of the training data. This paper aims to address the uncertainty surrounding whether the training data encompasses all test environments and waveform configurations and whether the trained model performs robustly in real-world environments. To this end, we propose a post-deployment joint measurement-recovery framework for AI-based transceivers, called VERITAS. VERITAS continuously detects changes in the distribution of received signals and triggers finite retraining. VERITAS monitors the wireless channel using 5G pilots and uses auxiliary neural networks to detect out-of-distribution channel profiles, transmitter speeds, and delay spreads. When such changes are detected, the conventional (reference) receiver operates in parallel with the AI-based receiver for a set period of time. Finally, VERITAS compares the bit probabilities of the AI-based and reference receivers for the same received data input to determine whether to initiate the retraining process.
