Federated learning for clients' data privacy assurance in food service industry

Authors: GORJI, HAMED - University Of North Dakota; SAEEDI, MAHDI - University Of North Dakota; ZADEH, HOSSEIN - Safetyspect Inc; HUSAIRIK, KAYLEE - University Of North Dakota; MOJTABA, SHAHABI - University Of North Dakota; Qin, Jianwei - Tony Qin; Chan, Diane; BAAEK, INSUCK - Orise Fellow; Kim, Moon; AKHBARDEH, ALIREZA - Safetyspect Inc; MACKINNON, NICHOLAS - Safetyspect Inc; VASEFI, FARTASH - Safetyspect Inc; TAVAKOLIAN, KOUHYAR - University Of North Dakota

Submitted to: Applied Sciences
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 8/14/2023
Publication Date: 8/17/2023
Citation: Gorji, H., Saeedi, M., Zadeh, H., Husairik, K., Mojtaba, S., Qin, J., Chan, D.E., Baaek, I., Kim, M.S., Akhbardeh, A., Mackinnon, N., Vasefi, F., Tavakolian, K. 2023. Federated learning for clients' data privacy assurance in food service industry. Applied Sciences. 13(16):9330. https://doi.org/10.3390/app13169330.
DOI: https://doi.org/10.3390/app13169330

Interpretive Summary: The food service industry faces the challenge of ensuring that service facilities are clean and free of foodborne pathogens that may be hosted by organic residues and biofilms and can lead to foodborne diseases, putting customers at risk and compromising the reputations of service providers. New fluorescence imaging technology empowered by state-of-the-art artificial intelligence algorithms may help address this food safety issue, but using such advanced technologies raises concerns about data privacy and possible leakage of service providers’ sensitive information. This study tested the use of a decentralized privacy-preserving method, called federated learning, to analyze fluorescence video data captured using a handheld imager at eight different food service facilities (“clients”) and then develop a detection model for identifying potentially contaminated areas as seen by the handheld imaging device. Conventional model development and testing methods would involve sharing, compiling, or centralizing the data for analysis, but federated learning does not require data sharing across clients or data centralization on a server. Instead, an iterative process is used with a central server broadcasting parameters for a global model to individual clients, the clients conducting local dataset training using those parameters and then returning new parameters to the central server, and the server then updating the global model with the returned parameters and then broadcasting the updated global model to the clients for the next round of local dataset training. The central server handling the global model has no access to private client data, and only receives parameters for updating the global model. This study used two deep learning models (MobileNetv3 and DeepLabv3+) to find contaminated areas on different high-risk contamination surfaces (such as doorknobs, garbage cans, oven and refrigerator door handles, chopping boards, and preparation tables) captured in fluorescence video image frames, a federated learning framework called FedML and an aggregation algorithm called Fedavg. The resulting model was tested on data from two new client facilities that were not included in original model development and testing, and achieved 95.83% and 94.94% accuracies in classifying clean and contamination frames, respectively. The results demonstrated that using federated learning combined with fluorescence imaging and deep learning algorithms can improve safety and cleanliness assurance while assuring client data privacy. In the future, use of such kinds of privacy-preserving methods with advanced detection technologies involving machine learning and artificial intelligence will benefit businesses and consumers with improved food and product safety while reducing risks of information leaks that can be economically expensive and damaging to business operations and consumer confidence.

Technical Abstract: The food service industry faces the challenge of ensuring that service facilities are clean and free of foodborne pathogens hosted by organic residues and biofilms. Such contamination can lead to foodborne diseases, putting customers at risk and compromising the reputations of service providers. New fluorescence technology empowered by state-of-the-art artificial intelligence algorithms may address this issue. However, using such advanced technologies raises concerns about data privacy and possible leakage of service providers’ sensitive information. In this study, we employed federated learning, a decentralized privacy-preserving technology, to address client data privacy issues. By using federated learning, there is no need for data sharing across clients or data centralization on a server. We used a new fluorescence imaging technology with two deep learning models, MobileNetv3 and DeepLabv3+, to identify and segment the contaminated areas on different equipment and surfaces. We used FedML as our federated learning framework and Fedavg as the aggregation algorithm. The model was trained and validated on data from eight clients and tested on two new clients' data. The model achieved 95.83% and 94.94% accuracies (F-scores of 96.15% and 95.61%) for classification between clean and contamination frames for the data from the two new clients and resulted in intersection over union (IoU) scores of 91.23% and 89.45% for segmentation of the contaminated areas. The results demonstrated that using federated learning combined with fluorescence imaging and deep learning algorithms can improve safety and cleanliness assurance while assuring client data privacy.