People

Fabio Pinelli

Associate Professor
fabio.pinelli@imtlucca.it

Google ScholarORCIDLinkedInLinkLink

About me

I am an Associate Professor of Computer Science within the SySMA research unit of IMT Lucca. 

Currently, I am Research Associate to HPC Lab at ISTI, CNR, Pisa  

I worked as a research scientist at IBM Research, Ireland, Dublin; then, I held senior data scientist positions at Tiscali , Cloud4Wi, and Vodafone. I received my PhD in Information Engineering from the University of Pisa in 2010, where I worked at KDD Lab, CNR, Pisa. During my PhD, I was a visiting researcher at Senseable City Lab at M.I.T., Cambridge, MA, US.

News

Security in Federated and Distributed Machine Learning and Artificial Intelligence Environments

Contact persons: Fabio Pinelli, Alessandro Betti

Curriculum: Software, System and Infrastructure Security

Funds: University

Additional benefits: Full board accommodation

Website: https://sysma.imtlucca.it 

Description

Modern technologies increasingly use federated learning, which trains machine learning models across decentralised devices without transferring data to a central server, thereby enhancing privacy. For example, the next-word predictions on Gboard for Android devices are generated using this approach.


However, Implementing federated and distributed machine learning systems has introduced new challenges and opportunities in the cybersecurity landscape. These systems enable collaboration among different nodes, allowing models to be trained on distributed data without centralising the data themselves. However, this decentralisation introduces potential security vulnerabilities that must be effectively addressed to ensure data integrity and confidentiality.


The objective of the thesis is to address the following challenges and goals:

- Vulnerability Analysis: Conduct a detailed analysis of existing vulnerabilities in federated and distributed machine learning systems, including privacy threats, model manipulation attacks, and potential data security breaches.

- Development of Defence Techniques: Design and develop new defence techniques to mitigate the identified vulnerabilities, using approaches such as homomorphic encryption, secure and robust aggregation methods, and other advanced methods. The effectiveness of these defence techniques is evaluated through a series of case studies and practical experiments. 

- Integration: This also requires integrating the proposed solutions into existing federated learning frameworks and scenarios correlating the theoretical and practical aspects of the identified problems. 

From Threat to Tool: Fine-Tuning LLMs to Combat Disinformation in Digital Media 

Contact person: Fabio Pinelli

Curriculum: Data Governance & Protection

Funds: MUR DM 630— scholarship co-funded by a research institution where the student will spend 6 to 18 months of the PhD.

Additional benefits: Full board accommodation

Research Institution: CNR-IIT

Research Institution Contact Person: Marinella Petrocchi

Website: https://sysma.imtlucca.it/, https://marinellapetrocchi.wixsite.com/mysite

Description

The proliferation of large language models (LLMs) such as GPT-3 and GPT-4 has revolutionized natural language processing and various applications ranging from automated customer service to advanced tools for information retrieval. However, their potential to spread disinformation poses significant challenges. This PhD project aims to explore the dual role of LLMs as both enablers and mitigators of disinformation. By examining the mechanisms through which LLMs generate, amplify, and disseminate false information, we seek to understand their impact on public discourse and trust in digital media. The research will employ a multi-disciplinary approach, integrating computational linguistics, machine learning, and social sciences, to analyze large datasets from social media and other digital platforms. Key objectives include identifying patterns in LLM-generated disinformation, evaluating the efficacy of current mitigation strategies, and developing new techniques to enhance model transparency and accountability.

The main outcome of this study is envisioned to be the development of a fine-tuned LLM specifically designed to detect and mitigate disinformation. This model will be trained on a curated dataset containing both accurate information and known disinformation. The fine-tuned LLM will be optimized to recognize and counteract false narratives.

Research Activities

My research interests are Data Mining and Machine Learning and their application in different domains. In my early scientific career, I have mainly focused on the development of Data Mining frameworks for spatiotemporal data to be applied to Urban Dynamics and Intelligent Transportation systems. In the most recent years, I have worked on machine learning pipelines for business and marketing problems.

Currently, I am working on:

Recent Publications

Selected Publications 

See my Google Scholar profile or my DBLP page for a full list of publications.

Patents