HE-OFT: Privacy-Preserving One-Shot Federated Fine-Tuning under Homomorphic Encryption
HE-OFT fine-tunes models once under multiparty CKKS so no participant receives the trained model.
HE-OFT is a one-shot federated fine-tuning protocol in which each client keeps a low-rank adapter, uploads one encrypted classifier-head displacement, and never receives the combined model. The server aggregates contributions under multiparty CKKS without decrypting them, and a client quorum returns only the predicted label. On four text-classification tasks and one vision task, accuracy is 61 to 79 percent, versus 20 to 48 percent for a client training alone, retaining 85 to 96 percent of a disclosed model's accuracy. A query takes 56.1 to 255.1 seconds on a GPU, and server-side level restoration cuts per-query traffic from up to 1.6 GiB to 13.5 MiB.
- Clients upload one encrypted head displacement; the server never decrypts it.
- No participant receives the trained model; queriers get only a label.
- Accuracy is 61 to 79 percent, retaining 85 to 96 percent of a disclosed model.
- GPU queries take 56 to 255 seconds and traffic falls to 13.5 MiB.
From the original article
1 min read at arxiv.orgMany organizations adapt large pretrained models to their own tasks by fine-tuning on private data. Several of these parties often hold data for the same task and wish to fine-tune a model together without pooling that data. Federated learning (FL) enables joint fine-tuning, but reconstruction attacks on shared intermediate values (the model or its gradients) remain a privacy risk.
By Halil İbrahim Kanpak · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.