OPTIMIST: Opportunistic Federated Transfer Learning for Mobile Devices Using Sequential Independent Subnetwork Training

This paper introduces OPTIMIST, a federated transfer learning framework designed for opportunistic settings, where stable networks are unavailable and centralized servers are impractical. OPTIMIST leverages pre-trained backbones and enables the collaborative training of classifier heads by resource-constrained, sensor-rich edge devices. In the absence of centralized orchestration, model parameters are exchanged between mobile nodes over transient connections. To avoid catastrophic interference resulting from varying node mobility and encounter patterns, OPTIMIST adapts a conflict-free strategy that splits the classifier head into smaller non-overlapping subnetworks. These subnetworks are trained locally, exchanged during encounters for incremental refinement, and seamlessly reassembled into a complete classifier head. We implemented a prototype for Android devices to evaluate its practicality in cross-device scenarios. Our experiments demonstrate its effectiveness in varying degrees of data heterogeneity. Additionally, under assumptions of relatively sparse data sets on edge devices, batch training time on the order of seconds supports the feasibility of this technique in real-world environments.