As edge and fog computing become central to distributed AI, there is growing demand for simulation tools that combine serverless orchestration with privacy-preserving federated learning (FL). Existing frameworks either model edge resources without FL awareness (e.g., iFogSim, FogFaaS) or support FL without simulating realistic edge constraints (e.g., Flower). We present FedFog, a modular simulator that unifies FL and Function-as-a-Service execution in heterogeneous edge–fog infrastructures. FedFog introduces a novel multi-objective scheduling policy that jointly optimizes latency, energy, and accuracy under cold start delays, client heterogeneity, and data drift. Privacy and robustness are enhanced through integrated differential privacy noise injection and robust aggregation (coordinate-wise median). The corrected FedAvg implementation ensures theoretical soundness, with convergence validated against baseline FL frameworks. Extensive experiments on EMNIST and HAR datasets, including large-scale simulations (up to 512 clients) and real-world Raspberry Pi deployments, show that FedFog reduces latency by 28%, energy consumption by 24%, and accuracy loss under 20% label-flipping from 10.5% to 3.8% compared to stateof- the-art baselines. FedFog provides a reproducible, extensible testbed for advancing scalable, secure, and resource-efficient FL in dynamic edge environments.

Keywords