March 20, 2023 at 2 p.m in amphi Gaston Planté.
New ROC Ph.D. students will present their research activities on March 20, 2023, followed by short presentations from three master interns.
Find more details about the talks below.
Quantum differential privacy on chip
Data is a powerful resource nowadays. It is used to provide statistics, to improve the process of decision making, or more, it is the essence of powerful machine learning models that we find in today’s applications. However, Data is often sensitive, it may reveal something personal about an individual and cause harm to his privacy. Therefore, Governments and institutions impose different laws for collecting, storing, processing and publishing data such as in General Data Protection Law (GDPR) in the EU.
To satisfy the conditions imposed by these laws, many traditional techniques have been widely used by companies. These traditional techniques consist of anonymizing the data and releasing statistics instead of raw data. Now, these techniques are deprecated because it has been proven that they cannot counter some well-known attacks like reconstruction attacks, linking attacks, membership inference and so on. This is why finding other techniques to protect privacy is still an active research area.
That’s where Differential privacy (DP) comes in. This revolutionary technique allows us to protect sensitive data while still extracting valuable insights from it. Big companies like Google, Microsoft, and Apple have already implemented differential privacy in their real-world applications. However, there are still many challenges and issues to be addressed in this field.
In this talk, we will see the limitations of traditional techniques as a motivation to DP. We focus on DP before we give the open challenges and directions. Then we explore how quantum computing on chips may be a key for solving some of these challenges in the context of federated learning.
Rezak Aziz is a PhD student at Cnam, where he is currently pursuing research in privacy preserving techniques. He graduated with a master’s and an engineering degree from the Higher National School of Computer Science in Algiers (ESI ex. INI) in September 2022. Before that, he was an intern at Le Cnam during the master’s internship, where he was working on privacy preserving machine learning using homomorphic encryption. Now, he is exploring differential privacy and quantum computing as tools for privacy preserving techniques in the context of machine learning.
Dynamic Graph Neural Networks for Intrusion Detection in IoT Networks
The Internet of Things (IoT) has increased the number of smart devices connected to the Internet, which poses security challenges, especially during the COVID-19 pandemic. To overcome this problem, Network Intrusion Detection Systems (NIDS) based on Graph Neural Networks (GNN) have been explored. Such systems examine the interactions between the nodes in the network and use the hidden structural dependencies to detect abnormal activities. However, most existing GNN-based NIDS train the model on a static representation of the network, which make them unable to handle the temporal evolution of the IoT network.
This limitation gives rise to the need for dynamic approaches such as dynamic GNN that incorporates temporal information into the models and adapts to changes in network topology over time for more effective intrusion detection.
Advanced Modelling of Network Intrusion Detection Systems in IoT using Graph Neural Networks
Ihab Abderrahmane DERDOUHA
The complexity and diversity of IoT devices presents a huge challenge to standard security measures, including Network Intrusion Detection Systems (NIDS). In fact, while deep learning-based NIDS are capable of detecting patterns and anomalies in large and complex datasets generated by IoT devices, they have limitations when it comes to exploiting hidden relationships within network flows. On the other hand, Graph Neural Networks (GNN) take into account the complex interactions between different nodes in the network, and thus, GNN-based NIDS are more effective in detecting malicious activities.
Nevertheless, some obstacles remain, such as the time and resources required to detect intrusions. Therefore, advanced NIDS architectures, such as distributed NIDS, are being explored, leveraging ideas such as transfer learning and federated learning to develop more effective security systems for IoT networks.
Self-Sovereign Identity pour les dispositifs IoT et Consensus
Self-Sovereign Identity is emerging as the new paradigm for digital identity, which will replace centralized systems. Digital identity is not exclusive to people and can therefore be extended to IoT devices that typically consist of micro controllers with low to medium computing power and a small amount of memory, which can own and manage their identities based on their life cycle. In general,IoT systems use a server-client communication model in which devices are identified, authenticated and connected through servers with high compute and storage resources. Blockchain technology can be a solution for implementing such a system.
The blockchain represents a distributed database in the form of a chain of blocks of information obtained through a process of consensus among the participants.
The challenge in this configuration is to choose the consensus algorithm in such a way that its complexity is adequate to the IoT environment and that takes into account the limitations of these devices.