Miloudi Amara

Federated learning has emerged as a promising paradigm for collaborative machine learning in healthcare, enabling medical institutions to jointly train models while preserving patient data privacy and complying with regulations like HIPAA. However, its decentralized nature makes it vulnerable to model poisoning attacks, where malicious participants can compromise the global model’s performance, potentially leading to incorrect diagnoses or treatment recommendations. In this paper, we propose a novel hybrid defense mechanism that combines statistical anomaly detection with random partial aggregation to protect against both obvious and subtle poisoning attacks in healthcare FL systems. Our approach first employs a robust anomaly detection system to identify and filter out clearly malicious updates, then applies a random partial aggregation strategy to minimize the impact of subtle poisoning attempts that evade detection. Unlike existing methods that focus solely on either detection or robust aggregation, our hybrid approach provides two layers of defense working in a complementary fashion. Experimental results on the MNIST dataset, serving as a proof of concept that can be extended to medical imaging and clinical data, demonstrate that our method maintains model accuracy while significantly reducing the impact of poisoning attacks compared to the traditional FedAvg approach. Our approach successfully detects obvious attacks through anomaly detection while mitigating subtle poisoning through random aggregation, achieving a better balance between model performance and security. The proposed method is computationally efficient and can be readily integrated into existing healthcare FL systems, offering a robust solution for secure collaborative learning in medical environments where model reliability and patient safety are paramount.

Blood cancer (leukemia) is one of the most serious diseases that affect blood-forming tissues. It usually involves white blood cells. The early detection of this severe disease helps doctors to provide efficient treatment. However, the discovery of this sickness at its first stages is often not easy due to similar morphological characteristics of malignant and healthy blood cells. Flow cytometry was the only used technique for early detection of leukemia, but it is very expensive and usually unavailable in hospitals. Deep learning is a machine learning technique that teaches computers to do what comes naturally to humans learn by examples. In the last few years, deep learning has achieved great successes to solve concrete problems. In particular, it has proven successful in medical imaging classification. In this work, we propose a Convolutional Neural Network (CNN) experiment for the classification of malignant white blood cells from normal ones using a dataset of microscopic images. The proposed approach leads to a balanced model that reaches a high level accuracy.