Privacy-Aware Split Neural Network for Vertical Federated Learning Gizlilik Duyarli Par ali Sinir Agi Modeli ile Dikey Federe grenme

ÖZTÜRK H., AKAY B.

2025 Innovations in Intelligent Systems and Applications Conference, ASYU 2025, Bursa, Türkiye, 10 - 12 Eylül 2025, (Tam Metin Bildiri) identifier

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Doi Numarası: 10.1109/asyu67174.2025.11208413
  • Basıldığı Şehir: Bursa
  • Basıldığı Ülke: Türkiye
  • Anahtar Kelimeler: Deep Learning, Split Learning (SL), Vertical Federated Learning (VFL)
  • Erciyes Üniversitesi Adresli: Evet

Özet

In this study, a learning architecture was developed that can be effectively applied in scenarios where data privacy and collaboration requirements are at the forefront. With this architecture, a Vertical SplitNN-based solution was developed by combining Vertical Federated Learning (VFL) and Split Learning (SL) approaches. Within the scope of this architecture, clients, each of which possesses different feature sets, train their own local models, while only the active client, having access to label information, undertakes the training of the global model. In this way, data privacy is preserved, communication costs are reduced, and the need for a central server is eliminated. The developed architecture was tested on three different real-world datasets: Breast Cancer Wisconsin Diagnostic, Pima Indians Diabetes and Diabetic Retinopathy Debrecen. In the Breast Cancer dataset, 98 % accuracy, 0.0555 loss, and 0.9583 F1 score were obtained, while in the Pima Indians Diabetes dataset, 80 % accuracy, 0.1610 loss, and 0.6875 F1 score were achieved. Additionally in Diabetic Retinopathy Debrecen dataset, 84 % accuracy, 0.1257 loss, and 0.8462 F1 score were obtained. As a result of these findings, this architecture can be effectively used in especially developing privacy-focused and reliable artificial intelligence solutions in the field of healthcare domain without compromising performance.