Secure and Explainable Healthcare with Federated learning


In regards to the Healthcare, different hospitals gather information every day. Collectively they have a large dataset on which deep learning models can be trained to help clinical practitioners. However, the health data is like digital footprint, which contains private information of the data owner, and this creates privacy concerns. Therefore the data owner (individual Hospital or the patients) are not willing to share data. Moreover, deep learning models are like black box, we do not really know the reason behind their predictions, but in healthcare, clinical practitioners should know reason for a particular predication.  Hence, in the frame of Secure and Explainable Healthcare with Federated learning (SEHFL), we with our UK based partner Shujun Li (University of Kent, UK) propose to develop an end-to-end smart healthcare framework for real-time use, based on the advanced Internet of Things (IoT), Artificial Intelligent (AI), Explainable AI and Blockchain. A general overview of the proposed framework is given in Figure 1.

 Federated learning allows data experts to design artificial intelligences (AI) without compromising user privacy. In other words, Federated learning is  a mechanism of training a shared global model with a central server while keeping all the sensitive data in local institutions where the data belong, provides great promise to connect the fragmented healthcare data sources with privacy-preservation. This method is designed to disrupt the paradigm of centralized AI, in which an algorithm improves as more and more personal data is collected. Federated learning could be the solution for industries where data cannot be transferred to third parties for confidentiality reasons (health, banking, etc.). In particular, the following are the main objectives of the project: (1) developing new e-Health monitoring methods (a key component of the proposed smart healthcare system), this is particularly the case on e-Health monitoring applications for chronic patients, where the need for data quality to ensure correct decision making is very important. Patients monitoring refers to a continuous observation of a patient’s condition (physiological and physical) traditionally performed by one or several body sensors. In particular, the main objectives of the project are as follows: (1) construction of an online physiological monitoring system for chronic patients (where the need for data quality to ensure correct decision-making is very important) to collect data on the patient’s state of health using “smart clothing” instead of wearable sensors (significant actions and decisions are based on data from these sensors (eg: remote diagnosis, consultations, hospitalization, etc.); 2) develop new explainable federated learning techniques for ECG monitoring system and detection of cardiac arrhythmias, human activity recognition system (key elements of the proposed intelligent health system. (3) introduce a security framework based on blockchain to fight against model poisoning attacks and privacy in the proposed federated learning architecture, and (4) design of an explainable AI based model to explain the results of deep learning model.

We divide the proposed framework virtually into two tiers: tier 1 and tier 2. Here, tier 1 is more associated with providing smart accurate healthcare and explainability, whereas tier 2 is concerned with providing privacy and security to the underlying tier 1 setting. Each of the settings is discussed as follows.


We propose a framework for smart healthcare, for instance, arrhythmias detection using ECG data. We will fully integrate the sensors into intelligent garments such as shirts, which will send real-time signals of body activity to the edge devices without causing any discomfort to the user. The proposed system will then collect the data from a subject and de-noise it using the proposed convolutional neural network base auto-encoder. Following de-noising, we apply a convolutional neural network-based classifier to classify the input signal. For example, in the case of ECG signals, we classify them into five classes of arrhythmias: Nonecotic beats (normal beat), Supraventricular ectopic beats, Ventricular ectopic beats, Fusion Beats, Unknown Beats. This classification helps diagnose health issues (arrhythmias) and can save human lives. Moreover, we will develop new Human Activity Recognition (HAR) methods for the local edge servers as a key part of the proposed smart healthcare system. They will be used to monitor the health status of patients and to detect medical emergencies or assist the patients at home, thus enabling them to live on their own with enhanced confidence and quality of life Furthermore, we also classify other body activities like walking, jogging and running, etc., because they can be key indicators of health status. 

The proposed framework in FL architecture works as follows: we will first train separate models at each “edge (Hospital)” server locally with the local data available and then verify these local and global trained models using the proposed POM (POM will be discussed later). After successful verification, these models will be sent to the healthcare cloud service to be combined into an updated master model. When the “edge” servers acquire more data, they can download the latest master model from the server, update it with the new data, and send it back to the server. Throughout the process, raw data is never exchanged—only the models, which cannot be reverse-engineered to reveal that data. The edge servers will be equipped with some machine learning models to allow intelligent local responses to the patient’s needs in real-time but will also be able to infer when requests and processed (e.g., anonymized) data should be sent to the remote cloud server for getting additional support from the more advanced machine learning models there. By doing so, we will have an Edge-Machine-Learning-based (EML-based) smart healthcare system to provide a personalized resource service (provided by the local edge servers) for patients, to save the storage space at the remote cloud server, and to improve privacy and security protection of patients’ data. Considering the constrained resources of edge devices, we will propose an efficient way of training the local machine learning models and the request inference module for handling high-density IoT data streams. 

Following the above, we will design an Explainable AI based module. Explainable AI is artificial intelligence (AI) in which the results of the solution can be understood by humans. It contrasts with the concept of the « black box » in machine learning where even its designers cannot explain why an AI arrived at a specific decision. Hence, we propose to join explainable AI based module at the top of every trained model to visualize the prediction made by the underlying classifier. Moreover, the module will be generalized enough to work with any type of machine learning model architecture. This gives more trust and confidence to the clinical practitioners about certain predictions of a machine learning model

Tier 2:

In tier 2, we propose to increase the security and privacy measure Federated setting in tier1, where the data (weights of the trained model) from the edge devices are sent to the global cloud serve. The global cloud server then aggregates the collected weights from the edges device and updates a global model with new aggregated weights. The edge device downloads the global model, which provides an efficient and reliable real-time classification of underlying health issues. Moreover, as only weights being share into the cloud from the edge devices, it provides data privacy and security to the end-users. Thus, our proposed model also complies with the data protection guidelines of the General Data Protection Regulation (GDPR).

However, in the above setting, a malicious global server can breach privacy and security. To solve these privacy and security concerns, we further proposed to use secure aggregation and Homomorphic encryption at the global cloud server. This provides privacy protection in case of a malicious global cloud server. Additionally, we also proposed to use state-of-the-art blockchain technology to introduce trustlessness, security against model poisoning attacks, and privacy into the proposed federated architecture. A blockchain is typically managed by a peer-to-peer network collectively adhering to a protocol for inter-node communication and validating new blocks with the help of a consensus algorithm.

We will propose a new efficient and secure consensus algorithm called proof of model (POM) that will make ensure that both the local models and the global model contribute constructively to increase the accuracy of the global updated model before they are being used to update the global model. The POM also removes the possible model poisoning attacks by verification of model.  This makes the proposed system, one of a unique kind to provide a trustless, secure, privacy-preserving, and distributed deep learning architecture for the smart healthcare system.

Partners :

  • Shujun Li (University of Kent, UK), school of computing ,

Gemtex Scientific Managers :

Aller au contenu principal