We invite You to the next open seminar as part of the series "Application of artificial intelligence methods in diagnostics and control".
The lecture entitled
"Modeling the threats of large language models"
will be held
🟡 When? 0️⃣6️⃣.0️⃣5️⃣.2️⃣0️⃣2️⃣6️⃣ at 13:30 - 15:00
🟡 Where? In the Room 36 at the Faculty of Mechanical Engineering, silesian University of Technology
The seminar will albo be available online via ZOOM:
🔵 The link: bit.ly/4h7AOlM
🔵 Meeting ID: 927 4351 6244
🔵 Access code: 479284
During the lecture, Dr. Adrian Kapczyński will present an in-depth analysis of the architecture of large language models from the perspective of identifying and classifying attack sizes. He will discuss a.o.:
✅ input manipulation techniques,
✅ data exfiltration mechanisms,
✅ adversary attacks affecting the integrity and predictability of models,
✅ risk of leakage of sensitive information,
✅ safety of RAG mechanisms,
✅ approaches to threat modeling for AI systems in relation to standards such as MITRE ATLAS.
🔵 Biography:
Adrian Kapczyński, PhD, Eng. – a student of Professor Andrzej Grzywak, passionate about computer science, professionally associated with computer science and its applications for 28 years. Teacher, manager, futurologist, internal auditor ISO 27001 and ISO 42001, lead auditor ISO 27001 and ISO 22301, strategic consultant and cybersecurity expert, specializing in biometrics and steganography.
Within the Silesian University of Technology, he is the coordinator of the research sub-area of Cybersecurity, the head of postgraduate studies "Cyber Science. Cybersecurity Management", the leader of the research team and the supervisor of scientific circles.
Chairman of Committee No. 309 (for biometrics) within the Polish Scientific Committee. Animator of the "17 53c" cybersecurity community. Member of ISACA, IEEE, ACM, ISSA and Mensa Polska.
🔵 Abstract of the lecture:
During the lecture, an in-depth analysis of the architecture of large language models (LLMs) from the perspective of attack surface identification and classification, covering the full life cycle of the model, will be presented.
Key categories of threats specific to generative systems will be discussed, including input manipulation techniques, data exfiltration mechanisms, and adversary attacks affecting the integrity and predictability of model operation.
In addition, issues related to the risk of leakage of sensitive information and the security of Retrieval-Augmented Generation (RAG) mechanisms in the context of privacy protection and data access management will be presented.
In addition, threat modelling approaches developed for AI-based systems will be presented, along with their reference to recognised standards such as MITRE ATLAS.
The lecture will focus on current research directions and practical implications of designing secure systems using large language models.
