sabri belmadoui

During my internship at BPCE-IT, I have been working as a Data Scientist Intern with two main objectives: Research & Development (R&D) and operational tasks (the metier side). - Research & Development: My primary research focus is on leveraging Large Language Models (LLMs) to analyze user and system behavior within a branch of BPCE-IT. Specifically, I am investigating the application of LLMs for tabular time-series anomaly detection. This involves identifying unusual patterns and behaviors that may indicate potential issues or threats. - Key Projects: Throughout my internship, I have contributed to several significant internal projects, including: * Anomalous Behavior Detection on "Signaux Faibles": Identifying weak signals that have triggered an alert but were classified as 'not a threat' by the static solution, but they were. * User Access Anomaly Detection: Detecting unusual patterns in user access to ensure security and compliance. * Non-Active VM Detection: Identifying non-active virtual machines (Idle) with little to no references/information. - Collaboration and Team Interaction: I have collaborated closely with various teams, including: Security (SOC), Observability, Authentication, and System Administration. Teamwork has been an integral part of my internship, providing me with opportunities to engage with diverse teams and gain insights into various aspects of the Security Operations Center (SOC) and beyond. - Methodologies and Technologies: * Agile Methodology: Employed Agile practices using Jira for project management and task tracking. * Splunk: Utilized for monitoring, searching, and analyzing machine-generated data. * PyOD, Sickit-learn, other ML packages: Used for outlier detection in multivariate data. * LLM (hugginface & pytorch): Involved in fine-tuning tasks, knowledge-based embeddings, and Masked Language Modeling.

This research project focused on modeling human decision-making strategies using Recurrent Neural Networks (RNNs). In collaboration with the ENSEA and ETIS research laboratory, I explored various RNN variants, including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Elman networks, to study four primary decision-making strategies employed by humans. - Key Contributions and Learnings: One of the main takeaways from this project was understanding the intricate inner workings of different RNNs. I developed custom code from scratch to modify one of three specific mechanisms in an RNN, effectively transforming it to behave like another model. This included: * Altering the GRU architecture by disabling the gating mechanism to mimic the behavior of an Elman network. * Modifying the LSTM network to function as a reservoir computing model. * and much more. These architectural modifications highlighted the importance and impact of specific mechanisms within RNNs that enhance their effectiveness in handling and adapting to decision-making sequences. - Skills and Technologies: * RNN Variants: LSTM, GRU, Elman Networks * Custom Code Development: Writing and altering RNN architectures from scratch (Pytorch). * Mechanism Analysis: Understanding and modifying gating and memory mechanisms in RNNs * Research Collaboration: Working closely with academic and research institutions to conduct and present findings.

Title: Explainability in deep model learning applied to IDS. - Developed an intrusion detection system (IDS) using ML and DL techniques and carried out a comparative study on the CIC-IDS-2017 dataset. - Processed, cleaned and improved the CIC-IDS-2017 dataset, using SMOTE technique to resolve class imbalance. - Used explainability techniques such as DALEX, SHAP and LIME to extract meaningful information, presenting it in an easily understandable visual format.

Title: Hand gesture recognition for the detection of Arabic sign language. - Created a web application for real-time recognition of Arabic sign language hand gestures using Deep Learning techniques. - Generated a dataset of 15,000 images representing a variety of Arabic Sign Language hand gestures, applying preprocessing and augmentation methods to improve data quality. - Carried out a comparative study with different DL models on the created dataset.