DarkDistill: Difficulty-Aligned Federated Early-Exit Network Training on Heterogeneous Devices

Early-exit networks (EENs), which adapt their computational depths based on input samples, are widely adopted to accelerate inference in edge computing applications. The effectiveness of EENs relies on difficulty-aware training, which tailors shallow exits for simple samples and deep exits for complex ones. However, existing difficulty-aware training schemes assume centralized environments with sufficient data, which become invalid with real-world edge devices. In this paper, we explore difficulty-aware training in a federated manner, where EENs are collaboratively trained on heterogeneous devices. We observe the cross-model exit unalignment phenomenon, a unique problem when aggregating local EENs into a cohesive global model. To address this problem, we design a novel Difficulty-Aligned Reverse Knowledge Distillation scheme named DarkDistill that preserves the difficulty-specific specialization for aggregating heterogeneous local models. Instead of direct parameter averaging, it trains difficulty-conditional data generators, and selectively transfers generated knowledge of specific difficulty among matched exits of heterogeneous EENs. Evaluations show that DarkDistill outperforms the state-of-the-arts in both full-parameter and parameter-efficient fine-tuning of EENs.