HASA: Heterogeneity-Aware Subnet Allocation | IEEE EDGE 2026 Best Student Paper

Amir Hossein Shahdadian presents HASA, a method for assigning different subnet sizes to clients in model-heterogeneous federated learning. In federated learning, clients can have very different data as well as different compute resources. Existing approaches usually decide model size mainly based on resource limits. HASA looks at another question: can we use the same compute budget more effectively by also considering how different each client’s data is? HASA measures the heterogeneity of each client’s training data and uses this information to assign subnet widths. Clients with stronger distributional differences can receive wider subnets, while the total compute budget remains fixed. On an article-title next-word prediction benchmark, averaged over 10 seeds, HASA improves mean client accuracy from 13.82% to 14.32%. Worst-client accuracy increases from 11.47% to 11.99%, and P10 accuracy improves from 12.04% to 12.52%. The results show that model capacity does not need to be distributed equally across clients. Under the same training budget, a better allocation can improve both average and lower-tail client performance. The presentation also includes a hands-on Colab demo that walks through the full experiment. The paper received the Best Student Paper Award at IEEE EDGE 2026 in Sydney, Australia. 📄 Paper: https://arxiv.org/abs/2606.07621 💻 GitHub Repository: https://github.com/Reliable-and-Secur... ▶️ Colab Demo: https://colab.research.google.com/git... 🔗 LinkedIn: https://lnkd.in/p/dT_yasub 🌐 RSS Lab: https://www.rss-lab.de Authors: Amir Hossein Shahdadian, Ahmed M. Abdelmoniem, Mahdi Taheri, Samira Nazari, and Christian Herglotz