CS 590/690 SEMİNERLERİ
1. Privacy-Preserving CNV Calling through Homomorphically Encrypted Federated Fine-Tuning of ECOLE
Ahmet Arda Ceylan
Master Student
(Supervisor:Asst.Prof.Ercüment Çiçek) Computer Engineering Department
Bilkent University
Abstract: Accurate detection of Copy Number Variants (CNVs) in Whole Exome Sequencing (WES) data is critical for diagnosing genetic disorders and cancer. While the ECOLE model establishes a new state-of-the-art by utilizing a transformer-based architecture to achieve high precision and recall, its clinical utility often depends on fine-tuning with small, expert curated datasets or specific tumor samples. However, the centralized collection of such highly sensitive genomic data poses significant privacy and regulatory hurdles. This presentation introduces a novel framework for privacy-preserving transfer learning on the ECOLE model. We propose a Federated Learning (FL) architecture that enables multi institutional collaboration without the exchange of raw patient data. To ensure cryptographic security during the aggregation of model updates, we utilize Homomorphic Encryption (HE). Our approach strategically freezes the core transformer blocks and performs HE enabled fine-tuning exclusively on the remaining layers.
DATE: March 09, Monday @ 15:30 Place: EA 502
2. Deep Large-Margin lp-SVDD with CNN Feature Learning for Novelty Detection
Alireza Dastmalchi Saei
Master Student
(Supervisor:Asst.Prof. Shervin R. Arashloo) Computer Engineering Department
Bilkent University
Abstract: Compared with other approaches, the recently proposed lp-norm large-margin Support Vector Data Description (SVDD) method has demonstrated superior performance for novelty detection across diverse evaluation settings. Nevertheless, because it is designed to operate on fixed features, it inevitably decouples feature extraction from classifier training, leading to suboptimal performance. In this study, we extend the large-margin lp-SVDD approach by optimizing the objective function while jointly learning deep convolutional features, thereby minimizing classification errors and improving overall performance. To this end, we propose a stable alternating optimization scheme that performs classifier boundary fitting via a Frank-Wolfe-based algorithm, while a robust primal-based strategy guides CNN updates. Extensive experiments on several widely used novelty detection benchmarks demonstrate the superiority of the proposed approach over both the baseline and recent methods, even when only limited amounts of anomaly training data are available.
DATE: March 09, Monday @ 15:50 Place: EA 502
3. Automated Classification of Software Bugs: Leveraging Large Language Models to Distinguish Bohrbugs and Mandelbugs
Mehmet Taha Demircan
Master Student
(Supervisor:Assoc.Prof.Eray Tüzün) Computer Engineering Department
Bilkent University
Abstract: Research on the categorization of software bugs into Bohrbugs and Mandelbugs often hinges on the ability to interpret complex, non-deterministic failure patterns. Recent advancements in reasoning-oriented LLMs offer a new paradigm for analyzing bug reports and system logs with higher-level semantic understanding. This study investigates the efficacy of reasoning LLMs in automated bug classification, specifically distinguishing between Bohrbugs and Mandelbugs. The research further extends into the sub-categorization of Mandelbugs into aging-related and non-aging-related types, as well as more granular groups based on specific system behaviors and failure mechanisms. By leveraging the inherent causal inference capabilities of these models, the proposed approach aims to capture the underlying root causes of software faults more effectively than static analysis or traditional classification techniques. The findings suggest that reasoning LLMs not only enhance classification performance across these complex taxonomies but also provide improved interpretability, representing a significant shift toward more sophisticated, automated debugging systems.
DATE: March 09, Monday @ 16:30 Place: EA 502