65th Vienna Deep Learning Meetup: LLM Security & Synthetic Data Generation

Dear Deep Learners,

Our March meetup takes place on March 19th, 2025 at Dynatrace. We'll have two exciting topics: LLM Security and High-Quality Synthetic Data Generation.

***
Agenda:
18:15

• Arrival & registration

18:30

• Introduction by the meetup organizers
• Welcome by the host company

18:45

• Talk 1: LLM Security Threats: Prompt Injection, Jailbreaking, and Protecting LLM Applications
  Muhamed Loshi, Raiffeisen Bank International

19:30

• Announcements
• Networking Break & Discussions

20:00

• Talk 2: TabularARGN: Fast, Flexible, and High-Quality Synthetic Data Generation for Real-World Applications
  Paul Tiwald, Head of AI Research, Mostly AI

21:00

• Networking & Discussions

22:00 Wrap up & End
***
Talk Details:

Talk 1: LLM Security Threats: Prompt Injection, Jailbreaking, and Protecting LLM Applications

Artificial Intelligence, particularly Large Language Models (LLMs), is becoming an integral part of applications in business, customer service, automation, and decision-making. However, with their growing adoption, security risks have also emerged. This presentation explores real-world security threats to LLM applications, including prompt injection attacks, jailbreaking techniques, and their potential consequences—ranging from data exfiltration to the hijacking of autonomous tools. We will also discuss high-level security measures recommended by industry experts, including insights from the OWASP AI Exchange, to mitigate these risks.

About the Speaker:
Muhamed Loshi, a cybersecurity expert with 10+ years in defensive and offensive security. His current role includes leading AI security activities at Raiffeisen Bank International, ensuring secure AI adoption across the RBI Group. Muhamed also contributes to OWASP AI Exchange and co-authored security requirements for the EU AI Act.

Talk 2: TabularARGN: Fast, Flexible, and High-Quality Synthetic Data Generation for Real-World Applications
by Paul Tiwald, Head of AI Research, Mostly AI

Tabular synthetic data is transforming the way we work with sensitive and complex datasets, offering a privacy-preserving alternative while maintaining statistical fidelity. In this talk, I will introduce the concept of tabular synthetic data, why it matters, and how different approaches tackle its generation. I will then present Tabular Auto-Regressive Generative Networks (TabularARGN), our framework that achieves state-of-the-art synthetic data quality while being orders of magnitude faster than existing methods.

Beyond performance, synthetic data enables a range of real-world applications beyond privacy protection. In the second part of the talk, I will explore practical use cases where synthetic data provides significant value—from machine learning model development to fairness adjustments, missing value imputation, and data augmentation for imbalanced datasets.

About the speaker:
Paul Tiwald is a trained physicist with a PhD in theoretical physics. In 2017, he joined MOSTLY AI, where he now leads the AI research team. His work focuses on advancing synthetic data quality and developing new synthetic data features that unlock real-world use cases.

The talks will be followed by networking and discussions.
We kindly thank Dynatrace to provide drinks & snacks at this meetup.

If you want to suggest topics, speakers or venues for future meetups, please contact us at contact(at)vdlm.at

Looking forward to another great event in our meetup series!
Your VDLM organizers.