A Modular Federated Learning Framework for Privacy-Aware Food Adulteration Detection

Abstract

Adulteration of food remains a significant threat to public health, economic justice, and confidence in global food supply chains. Traditional detection methods, including laboratory physical and chemical testing and centralized machine learning, face several limitations such as high cost, latency, limited accessibility, and privacy concerns. To overcome these challenges, we propose a novel, modular and scalable federated learning framework tailored for food fraud detection. The system employs an autoencoder to compress data into a lower-dimensional space and reconstruct it, enabling efficient representation learning and anomaly detection. Unlike conventional methods, the framework ensures data privacy by keeping data local while collaboratively training models across distributed, heterogeneous datasets. It also extends beyond binary classification to capture complex, realworld adulteration patterns. The results demonstrate significant improvements of our federated model against various baselines, achieving more than $30 \%$ higher precision and more than $40 \%$ higher recall, thus improving accuracy, robustness, and scalability for food adulteration detection.