Keynote Lectures

Rethinking Privacy Protection in Federated Learning in the Face of Model Inversion Attacks
Wenjing Lou, Virginia Tech, United States, United States

Securing Ultra-Large Scale Infrastructures: Challenges and Opportunities
Awais Rashid, University of Bristol, United Kingdom, United Kingdom

Turing's Echo on Deceptive Machines: The Challenge of Distinguishing Human and AI Creations
Ahmad-Reza Sadeghi, Technical University of Darmstadt, Germany, Germany

Abstract
The current success of machine learning has largely depended on centralized learning, which pools data from multiple sources to a central location. This presents significant challenges in domains like healthcare where patient data is often siloed across multiple institutions, and strict privacy regulations prevent centralized data sharing. Federated learning, a distributed learning paradigm allowing institutions to collaboratively train models without moving data across institutional boundaries, is thus highly advantageous due to its ability to maintain data locality and address legal and ethical barriers to data sharing. However, recent research has shown that federated learning is susceptible to privacy attacks, such as data reconstruction and membership inference, where sensitive information can be inferred from model updates.
In this talk, we will explore privacy challenges in federated learning by introducing a sophisticated model inversion attack called scale-MIA. This attack efficiently reconstructs clients’ training samples from aggregated model updates in federated learning and undermines the effectiveness of secure aggregation protocols. We will also discuss the impact of such attacks and explore emerging solutions to enhance privacy in federated learning systems.

Abstract
Digital infrastructures are seeing convergence and connectivity at unprecedented scale. This is true for both current critical national infrastructures, such as water, power, and emerging future systems that are highly cyber-physical in nature with complex intersections between humans and technologies, e.g., smart cities, intelligent transportation, high-value manufacturing and Industry 4.0. Diverse legacy and non-legacy software systems underpinned by heterogeneous hardware compose on-the-fly to deliver services to millions of users with varying requirements and unpredictable actions. This complexity is compounded by intricate and complicated supply-chains with many digital assets and services outsourced to third parties. The reality is that, at any particular point in time, there will be untrusted, partially-trusted or compromised elements across the infrastructure. This poses a range of fundamental questions: How does one measure the security state of such infrastructures? What are the complexities of managing security in a landscape shaped by the often competing demands of a variety of stakeholders? How does one secure infrastructures of such complexity or conduct incident response in such ultra-large-scale settings? In this keynote, I will discuss insights from a multi-year programme of research investigating these issues and the challenges to addressing them.

Abstract
As generative AI models evolve, distinguishing between human-generated and AI-generated content is becoming increasingly challenging, threatening trust across various domains such as misinformation in media, political campaigns, legal accountability, scientific integrity, and cybersecurity. Distinguishing between machine and human outputs will become vital because, in the dystopian future, machines will potentially rise against humans in various ways.
This talk explores methods and technologies for identifying the origin of content, focusing on audio and text. We briefly review the existing models and their limitations in detecting subtle differences between human-generated and AI-generated content. Our approach integrates physical principles, such as the Micro Doppler Effect, into machine learning frameworks, enriching models with prior knowledge and reducing bias in detection. We conclude by outlining key challenges and future directions in this rapidly evolving domain.