About this Abstract |
| Meeting |
2025 Annual International Solid Freeform Fabrication Symposium (SFF Symp 2025)
|
| Symposium
|
2025 Annual International Solid Freeform Fabrication Symposium (SFF Symp 2025)
|
| Presentation Title |
Federated Learning for Defect Detection in Additive Manufacturing: Mitigating Label-Flipping Attacks in Distributed Factories |
| Author(s) |
Md Sazol Ahmmed, Sriram Praneeth Isanaka, Sung-Heng Wu , Atiqur Rahman, Muhammad Arif Mahmood, Frank Liou |
| On-Site Speaker (Planned) |
Muhammad Arif Mahmood |
| Abstract Scope |
In modern manufacturing, Distributed Digital Factories (DDFs) are a revolutionary concept that enables collaborative manufacturing but face challenges in centralized quality control due to data privacy concerns. We propose an FL framework enhanced with MUD-HoG (Multi-Dimensional History of Gradients and Hierarchical Clustering), which detects malicious clients by analyzing gradient patterns without accessing raw data. To evaluate its effectiveness, Additive manufacturing data with six input features and three output classes have been used; we simulated a strong label flipping attack on the weakest client. As a result, the attack reduced global model accuracy to 35.07%, and class-wise F1-scores fell below 0.40. After applying MUD-HoG, the system successfully detected and excluded the compromised client, recovering global accuracy to 98.05%. The “Normal” class F1-score improved from 0.35 to 0.97, with over 25% gains in precision and recall for other defect classes. This method advances the deployment of reliable, privacy-preserving smart manufacturing networks. |
| Proceedings Inclusion? |
Planned: Post-meeting proceedings |