Federated learning (FL) enables multiple clients to train a shared model without pooling raw data. In practice, however, clients are not uniform: some hold harder data, and others have tighter compute budgets. This paper studies local early-exit inference for Vision Transformers (ViTs) as a way to handle such heterogeneity.
The title points to a "client-aware" and "security adaptive" design, suggesting that the early-exit strategy is adjusted based on each client's data difficulty, compute limits, and security needs. This is a practical direction for deploying ViTs in real-world FL settings where one-size-fits-all inference is inefficient.
The provided abstract is cut off after the opening sentence, so the specific mechanisms, experimental setup, and security model are not described in the source. This digest therefore reflects only the stated motivation and the general research direction. Where the source would offer further detail, it is not available here.