New Solution by University of Tehran Researchers to Enhance AI Security in Brain MRI Analysis
According to the Public Relations Office of the University of Tehran, in a study published in the journal Scientific Reports, University of Tehran researchers designed a framework called OptiGradTrust, developed to increase the resilience and trustworthiness of AI in brain MRI image analysis.
Federated learning is a new AI method that allows different centers, including medical and research centers, to participate in training a shared AI model without transferring patients’ raw data. In this method, primary data remain local, and only information resulting from training and model updates is exchanged between centers and the central system. This feature can help preserve patient privacy.
However, this method faces challenges. Differences in imaging devices, imaging protocols, patient characteristics, and the amount and quality of data available at each center can cause substantial variation among data from different centers. On the other hand, a center may introduce inappropriate information into the model training process due to technical errors, low data quality, or even intentional behavior. This can reduce the final model’s performance and is especially important in sensitive medical applications.
To address this issue, researchers introduced the OptiGradTrust framework. Rather than trusting information sent by each center equally, the framework evaluates it from six different aspects. Indicators used in this assessment include the degree of similarity between training information and the expected pattern, its consistency with information from other centers, the magnitude of changes made to the model, the compatibility of the information, and each center’s contribution to improving model performance.
Next, a mechanism based on reinforcement learning and attention networks combines the results of these evaluations to determine the level of trust in information sent by each center. As a result, information from centers with abnormal or suspicious behavior has less influence on the final model. The framework also seeks to maintain stability in the training process using an optimization method designed for heterogeneous and malicious conditions.
To evaluate the proposed framework, researchers tested it on several datasets, including MNIST, CIFAR-10, an Alzheimer’s disease MRI dataset, and the OASIS clinical dataset.
The results showed that the method performs well even when data from different centers differ. In the Alzheimer’s MRI dataset, under heterogeneous data conditions, the proposed framework achieved 95% accuracy, while compared methods such as FLGuard, FLTrust, and FLAME achieved lower accuracies.
One of the key aspects of this research is the combination of two important issues in medical AI: privacy preservation and increased model security. In the proposed framework, patients’ raw MRI data are not made available to other centers, and centers participate only in the process of training a shared model. At the same time, the system attempts to identify abnormal or malicious information before it seriously affects the model and to reduce its weight.
Despite the results, the authors emphasized that this research does not yet mean the method is ready for clinical use. In particular, the number of samples used in the OASIS clinical evaluation was limited, and the results of this part are considered only as proof of concept. Therefore, for a final assessment of the framework’s performance, testing it on larger, multi-center datasets such as ADNI and UK Biobank, as well as examining more complex attacks, is necessary.
The paper, titled “Byzantine robust federated learning for heterogeneous brain MRI using multisignal gradient fingerprinting and adaptive trust aggregation”, was published in Scientific Reports in collaboration with Mohammad Karami, a master’s student at the Faculty of Electrical and Computer Engineering; Dr. Fatemeh Ghassemi; and Dr. Hamed Kobriaei, faculty members at the Faculty of Electrical and Computer Engineering, University of Tehran.
The article is available at the following link:
Your Comment :