FedContrast: A Contrastive Learning Framework to Mitigate Client Drift Under Statistical Heterogeneity in Federated Multi-Class Soil Classification
Year: 2025
Type: preprint
Abstract: FedContrast is a novel contrastive learning framework designed to improve federated classification performance under statistical heterogeneity through representation alignment techniques. It was implemented on two complementary soil image datasets: JPSID (1,563 samples, 4 soil types) and PSST (163 samples, 5 soil types), collected from agricultural and environmental monitoring contexts. Classification performance was evaluated under both IID and Dirichlet-induced non-IID settings, with concentration parameters \(\alpha = 0.5\) and \(\alpha = 1.0\) , across client configurations of \(K = 5\) and \(K = 10\) . For non-IID setups, the InceptionV3 model was selected with a batch size \(B = 32\) and communication rounds \(T = 100\) , as it consistently outperformed VGG19, Xception, and MobileNetV2 even under IID conditions. Evaluation results show that FedContrast outperformed traditional FL methods, achieving 88.8% accuracy, 88.8% F1-score, 89.9% precision, and 87.7% recall on the JPSID dataset under extreme heterogeneity ($K = 10$, $\alpha = 0.5$). On the PSST dataset, FedContrast maintained superior performance with a total loss ( \(\mathcal{L}_\text{total}\) ) of 0.367 ($K = 5$, $\alpha = 1.0$) and 0.338 ($K = 10$, $\alpha = 1.0$), reflecting a 15--45% improvement over traditional FL methods such as FedAvg, MOON, FedSGD, and FedProx.
Source: Research Square
Authors: Aviral Kumar Goyal, Manish Pandey, Dhirendra Pratap Singh, Jaytrilok Choudhary, Rahul Haripriya
Institution: Maulana Azad National Institute of Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4411190873