Adaptive federated data pipeline optimization for low-latency AI inference in distributed edge computing environments

Abstract

Federated edge frameworks assume static pipelines, but real latency comes from ingestion, preprocessing, and aggregation delays. To address this shortcoming in federated edge intelligence, we propose Adaptive Federated Pipeline Optimization (AFPO). The framework jointly optimizes pipeline-stage allocation, federated aggregation, and model compression under dynamic edge conditions. We present a 3-layer architecture (sensing tier, federation tier, inference tier) that uses a Reinforcement Learning-based scheduler to assign pipeline stages based on real-time resource telemetry, a latency-aware aggregation mechanism that reduces the effects of stragglers, and an Adaptive Compression Module that dynamically adapts quantization and pruning based on latency feedback. We demonstrate that this problem is strongly NP-hard and establish that AFPO will converge under non-IID data, time-varying network conditions, and deadline-based straggler dropping. AFPO was evaluated empirically on the CIFAR-10 and Fashion-MNIST image-classification benchmarks. Compared to the baselines of FedAvg and FedProx, which are tailored to training, it cuts down inference latency by 35.7%, decreases the communication overhead by 40.6%, and boosts classification accuracy by 5.08 percentage points. This accuracy advantage is specific to those baselines. Against inference-optimized baselines reconstructed in the same simulation environment, AFPO attains the lowest latency and communication cost but slightly lower accuracy (0.896 versus 0.905–0.910). It therefore trades a small amount of accuracy for improved latency and communication efficiency. The RL-based scheduler is trained using a policy-gradient approach. The latency and communication-efficiency enhancements over rule-based, resource-aware, partition-based, and early-exit scheduling are statistically significant across eight random seeds with 95% confidence intervals. Significance applies to three non-IID severities, two network regimes and a second synthetic workload profile. For the image-classification workloads evaluated here, AFPO offers an effective approach to latency-aware federated edge intelligence in Internet of Things and autonomous systems applications.

The authors' abstract, as published at the source. Discover Computing, 2026 · DOI ↗