ADEx: Adaptive Early Exit DNNs for Inference Robustness to Popularity Drift

Deep neural networks (DNNs) with Early Exits (EE) are widely adopted at the network edge to reduce inference latency and resource consumption by allowing samples to exit the DNN early when sufficient confidence is achieved. However, their effectiveness significantly degrades under popularity drift – where the frequency of target classes shifts over time. This paper introduces ADEx, an Adaptive Drift-aware EE framework that enables dynamic and lightweight adaptation of DNNs with EEs to class popularity drift. ADEx continuously monitors EE behavior to detect popularity shifts and selectively retrains only the affected ones, while keeping the backbone network unchanged. For unsupervised adaptation, ADEx uses the final exit as a teacher to pseudo-label new data, applying a priority- aware loss that enforces high confidence on popular classes and uncertainty on others. Furthermore, we analyze execution strate-gies balancing adaptation speed, latency, and resource usage. Experimental results demonstrate that ADEx restores inference efficiency, reducing main-branch usage from a post-drift peak of 46.14% to 7.65% (near the 7.46% pre-drift level) and lowering mean latency from 3.05 ms to 2.57 ms. Compared to a Joint Fine- tuning baseline, ADEx achieves similar accuracy recovery while reducing peak GPU memory by 52.3% and avoiding full-model shadow copies.