The Spring Event (SE) is a space where first and second-year EMAI students will meet each other, share their experiences and build the community of EMAI students.

   

Second-year students will showcase and discuss their internship experiences, master's theses, and other research projects. First-year students will learn from senior peers and practice their debating skills by commenting on and discussing their colleagues' experiences.

Location

Programme

Wednesday, March 26, 2025
9.45 Welcome and opening remarks
10.00 Session 1:
  • Stjepan Picek. On the (In)Security of Deep Learning.
  • Lejla Batina. Reverse engineering neural networks through physical side channels.
11.00 Coffee break (mentoring, 2nd-semester students currently at Radboud University only)
11.30 First and Second-year students: Presentations (session 1)
12.30 Lunch break
13.30 First and Second-year students: Poster session (open to RU students)
15.00 Coffee break (mentoring, 2nd-semester students currently at the University of Ljubljana only)
15.30 Invited talk: Frederik Zuiderveen Borgesius [slides]
17.00 Internal meeting (only project coordinators)
19.00 Welcome dinner in De Hemel [link]

 

Thursday, March 27, 2025
8.30 Visit to NXP - bus trip [link]
15.00 Departure to Boat trip [link]

 

Friday, March 28, 2025
9:00 First and Second-year students: Presentations (session 2)
10:00 Invited Talk: Lucia Cavallaro. Embedded ML: progress and prospects
10.45 Coffee break (mentoring with 2nd-semester students currently at Sapienza University only)
11.00 Hands-on Session: Greg Alpar. Writing scientific paragraphs (laptop needed)
12.30 Lunch break
13:30 First and Second-year students: Presentations (session 3)
14.30 Session 2:
  • Vicenç Gómez. Recommender Systems and Society: The Unintended Consequences of Human-AI Coevolution
15.00 Town Hall (open to all the EMAI community) 
15.30 Coffee break
16.00 Pub Quiz: Pieter Montè.
Huygensgebouw, Gigabyte restaurant (we depart from Room 02.610 at 15:45) [link]
17.30 Farewell Dinner

 

  • Cristian Bassotto.  AI-based attacks on Post-Quantum Cryptography [slides]
    Quantum computers threaten cryptographic security based on hard problems like integer factorization. As quantum technology advances, systems may become vulnerable to decryption via Shor’s algorithm. Post-quantum cryptography (PQC) aims to counter this but faces challenges from AI-driven cryptanalysis. This thesis explores AI techniques, particularly deep learning, to assess the security of PQC algorithms like CRYSTALS-Kyber and develop benchmarks to resist quantum and AI-powered attacks.
     
  • Giuseppe Cianci. Multimodal Machine Learning for Risk Stratification of Patients with Atrial Fibrillation

    Atrial fibrillation (AF, heart arhythmia), can be treated by catheter ablation. For it to be successful, low voltage areas (LVAs) in the atrium need to be identified. However, intracardiac voltage mapping is an invasive procedure. Thus it is resource-intensive and has procedural risks. A non-invasive AI approach to estimate LVAs could transform AF management by reducing the need for invasive procedures. This thesis presents a multimodal ML model to predict LVAs using ECGs and other data.

  • Nazanin Niayesh. LLM Red Teaming
    The increased use of LLMs and their integration into many critical systems raises concerns about their security. Exploitation of vulnerabilities in these models through adversarial techniques such as prompt injection or jailbreaking can cause significant consequences for the systems and applications that rely on these models as well as their users. This project explores this emerging issue, focusing on how different types of prompt injection attacks may affect LLM models.

  • Honor Duthie. Monocular SLAM in Endoscopic Procedures [slides]
    In order to improve diagnostic outcomes in procedures such as colonoscopy we attempt to build a 3D reconstruction of the internal structure and texture of the surgical space, in real time, using only the video footage of the single endoscopic camera. How can we tailor the feature detection of the Simultaneous Localisation and Mapping (SLAM) pipeline to biological images in order to improve accuracy? 

  • Adrian Gheorghiu. Machine learning for radar sensing in advanced driver assisted systems [slides]
    Radar is a core perception technology in advanced driver-assisted systems (ADAS) that enhances safety and convenience. Future vehicle radar systems will likely move to the higher millimeter-wave bands and incorporate more digital processing and machine learning capabilities. Recent papers have already explored efficient automotive radar machine-learning techniques using raw range-azimuth maps and methods to use teacher-student training with privileged LLMs.

  • Ahmet Çalış. AI as a scientific assistant
    Given the overwhelming amount of new papers every day, new strategies are needed to find, categorize and understand them. In this project we will refine an existing mutliagents pipeline that searches, downloads and processes scientific papers to generate insights for scientific researchers in the European Commission using LLMs and graph techniques.

  • Neus Vegas Morales. Identifying key T-Cell Receptors for cancer prediction [slides]
    I will introduce T-cell receptors (TCRs) and their role in cancer prediction. I will discuss different ways to represent TCR data and explain how the Set Covering Machine algorithm can help identify a small but meaningful set of TCRs that improve prediction accuracy.
     
  • Bernardo Perrone De Menezes Bulcao Ribeiro and Camile Lendering. EdgeEar: Efficient and Accurate Ear Recognition for Edge Devices [slides]
    Deploying high-performing ear recognition models on resource-constrained devices is a challenging problem, limiting their applicability. In this presentation we introduce EdgeEar, a lightweight model based on a hybrid CNN-transformer architecture. By incorporating low-rank approximations into specific linear layers, EdgeEar reduces its parameter count by a factor of 50 compared to the current state-of-the-art, bringing it below two million while maintaining competitive accuracy.
     
  • Roberta Chissich. Field segmentation [slides]
    Agricultural fields are always changing whether it’s due to seasonal effects or activities like harvesting or ploughing. The outlines of the fields are an important input parameter for the algorithms to filter agricultural events. Within the field of delineation important advancements have recently been made and additional datasets have been published. The goal of this research is to develop a generalized ML model based on the latest insights that can find the field outlines globally based on high resolution optical satellite imagery and/or low-resolution multispectral satellite imagery.
     
  • Andreas Pattichis. Scalable AI for Infrastructure Mapping: Adapting School Detection to Open-Source Imagery and Expanding to Health Centers [slides]
    This research aims to adapt the Giga initiative's AI pipeline for school detection from high-resolution satellite imagery to open-source, lower-resolution alternatives like Mapbox. It focuses on enhancing scalability and cost-efficiency while maintaining detection accuracy.  Additionally, the project seeks to generalize the pipeline to identify other critical infrastructures, such as health centers, supporting universal connectivity mapping for education and healthcare services.
  • Aleksandar Dimitrievikj. Steering Vector-Based Molecular Generation Using Language Models [slides]
    Steering vectors are vectors added to the output of different layers of a NN that guide the model to generate responses that have a certain style, bias or characteristic. This project investigates whether steering vectors can effectively guide generative language models to produce molecules with distinct, user-defined properties. Our approach aims to improve targeted drug discovery by enabling fine-tuned molecular property optimization. 
     
  • Leonida Lumburovska. Thrombus Segmentation for Predicting Anticoagulant Efficacy in Portal Vein Thrombosis [slides]
    The project's main idea was to investigate deep learning-based liver, portal vein and thrombus segmentation in portal vein thrombosis patients using AI-based models such as nnU-Net, TotalSegmentator, and ResNet. Our approach aimed to improve segmentation accuracy and explore its potential for predicting canalization success with anticoagulants. Challenges included limited data and model adaptation.
     
  • Antonio Carpes. Robotic Autonomous Learning
    This thesis explores the integration of Imitation Learning (IL) and Reinforcement Learning (RL) to make robot task learning more efficient. By taking advantage of IL, robots acquire initial policies from proper demonstrations, which are then refined through RL using priors. These priors guide the learning process, preventing robots from relearning fundamental behaviors. Starting in simulation, our work will be tested in real robots, Franka and Tiago.
     
  • Nikolay Kormushev and Manfred Gonzalez Hernandez. Masking identities: Customizable Face Anonymization with Diffusion Models [slides]
    Diffusion models have demonstrated high-quality image synthesis and restoration results by merging learned distribution with input images. In this paper, we exploit their capabilities and propose a novel deidentification approach to preserve facial privacy at different levels by modifying the guidance of the diffusion process to blend input images with similar images featuring different identities. The results effectively balance a trade-off: generating high-quality images and preserving privacy.
     
  • Djibril Coulybaly. Landing a Quadcopter Autonomously on a Moving Platform
    With the use of computer vision, control theory and pose estimation, we explore the theory, technical implementation and practical applications of precisely landing a quadcopter autonomously on a moving platform.
  • Stjepan Picek. On the (In)Security of Deep Learning [slides]
    Deep learning achieves outstanding results in many tasks, ranging from computer vision and financial data to self-driving cars and robotics. Still, such algorithms are vulnerable to attacks like evasion, poisoning, and model stealing attacks. This talk will cover several attacks and defenses. We will start with a brief introduction to the security of machine learning, followed by recent results. We will finish with an overview of some open questions and challenges.
     
  • Lejla Batina. Reverse engineering neural networks through physical side channels [slides]
    Side-channel analysis has changed the field of cryptography and security and it became the most common cause of real-world security applications failing today. In this talk we first survey several examples of AI assisting physical attackers and discuss the impact of AI on the field of hardware security. We describe the way side-channel analysis can assist the AI model stealing by reverse engineering commercial neural nets architectures.
     

  • Vicenç Gómez. Recommender Systems and Society: The Unintended Consequences of Human-AI Coevolution [slides]
    Algorithmic ranking systems shape online interactions, influencing engagement, misinformation, and polarization. This talk explores the unintended consequences of popularity-based rankings, drawing on two studies. The first examines how engagement-driven algorithms fuel polarization and ideological extremism. The second reveals a "few-get-richer" effect, where ranking amplifies minority signals. Together, these insights highlight the profound societal impact of human-AI feedback loops.

How to reach the venue