Graduate School of Engineering and Science
Permanent URI for this collectionhttps://hdl.handle.net/11693/115678
Browse
Recent Submissions
Item Open Access Optical spectroscopy of interlayer excitons in the infrared region(2026-04) Ölçer, Ahmet OsmanVan der Waals heterostructures formed from transition metal dichalcogenide (TMDC) monolayers provide a versatile platform for studying interlayer excitons, where electrons and holes are confined in different layers. Owing to their long spontaneous emission lifetimes, tunable photoluminescence (PL) emission, permanent out-of-plane electric dipole moment, and favorable electronic properties, these IXs are promising for optoelectronics, valleytronics, and quantum photonic applications. In particular, MoS2/WSe2 heterostructures are promising for these because of interlayer excitons with emission in the infrared region close to the O-band. This thesis investigates the optical properties of infrared interlayer exciton emission in MoS2/WSe2 heterostructures by means of optical spectroscopy under various experimental conditions. Both MoS2/WSe2 heterobilayers and WSe2/MoS2/WSe2 heterotrilayers were fabricated and characterized. The experimental work includes the fabrication of the heterostructures and the development of microscope-based spectroscopy platforms for both room-temperature and cryogenic measurements. Using these systems, steady-state photoluminescence, timeresolved photoluminescence, temperature-dependent spectroscopy, and magnetophotoluminescence experiments were performed. The results reveal clear infrared interlayer exciton emission from the fabricated heterostructures and show how their spectral position, intensity, lifetime, and magnetic response evolve with various experimental conditions. A comparison between bilayer and trilayer regions demonstrates that the addition of the third layer to form symmetric heterostructures significantly modifies the excitonic properties. In particular, differences in emission wavelength, lifetime, circular polarization, and Zeeman splitting indicate that trilayer stacking leads to distinct interlayer exciton properties compared to bilayers. These findings provide insight into the role of layer arrangement and interlayer coupling in determining the properties of infrared interlayer excitons. Therefore, this work contributes to the understanding of excitonic processes in MoS2/WSe2 heterostructures and supports their potential use in infrared photonics, valleytronic platforms, and future quantum and telecommunication-related technologies.Item Embargo On the improved average distribution of the ternary divisor function in prime arithmetic progressions and the Möbius function over Beatty sequences(2026-04) Boran, MuhammetThis thesis studies distribution problems for arithmetic functions, focusing on the ternary divisor function d3(n) in arithmetic progressions and the Möbius function µ(n) along Beatty sequences. We say that d3(n) has exponent of distribution θ if, for every ε > 0, the expected asymptotic formula holds uniformly for all moduli q ⩽ x θ−ε . Using the Petrow–Young [1] subconvexity bound for Dirichlet L-functions, we strengthen this averaged result and obtain an exponent of distribution 8/11 [2] when averaging over reduced residue classes modulo a prime modulus q. We investigate sums of arithmetic functions over Beatty sequences Bα,β. We prove a nontrivial bound for µ(n) along Beatty sequences: for almost all α > 1, and we also discuss corresponding estimates for fixed parameters (α, β) under suitable Diophantine approximation hypotheses.Item Embargo Thermal drawing of fiber-based hybrid nanogenerators for integrated energy harvesting and sensing(2026-04) Hussain, Seyd ZubairWearable electronics have advanced rapidly in recent years and have been used for continuous, non-invasive monitoring of human health and activity through smart textiles. For practical and long-term use, such systems require sustainable power sources that eliminate reliance on external batteries while maintaining flexibility and comfort. Human motion, such as walking, jogging, and running, is a significant source of ambient mechanical energy, and harnessing it could provide a sustainable energy solution. Nanogenerators are a promising platform for converting this energy into electrical energy, particularly through piezoelectric and triboelectric effects. However, relying solely on these individual mechanisms is often insufficient for wearable electronics, making hybridization an effective approach for enhanced energy generation. Piezoelectric-triboelectric hybrid nanogenerators (PT-HNGs) offer an effective solution by efficiently converting low-frequency human motion into electrical energy for self-powered sensing. Here, we report a fiber-based platform that enables effective piezoelectric-triboelectric coupling within a single structure. The use of a thermal drawing fabrication technique allows scalable and continuous fiber production. The hybrid fiber integrates a graphene nanoplatelet (GNP)-reinforced poly(vinylidene fluoride) (PVDF) active layer, conductive polyethylene (cPE) polymer electrodes, and an engineered internal air gap between the polycarbonate (PC) and PVDF layers, which enhances electromechanical performance while maintaining textile compatibility. At an optimal GNP loading of 5 wt%, the device exhibits a 118% increase in open-circuit voltage and a 151% increase in short-circuit current compared to pristine PVDF fibers, delivering a peak power output of 21 µW and a power density of 266.27 mW m−2 . The PT-HNG enables stable, self-powered monitoring of physiological signals, including respiration, arterial pulse, and muscle activity, demonstrating its potential for next-generation smart textiles and wearable health monitoring. In addition, the device shows good mechanical durability and stable performance under repeated mechanical loading, highlighting its suitability for long-term wearable applications. The scalable fiber architecture further provides a practical pathway for integration into textile systems and real-world self-powered sensing platforms.Item Embargo Neural correlates of topographic reorganization in adult V1 following short-term monocular deprivation and perceptual learning(2026-03) Karahan, TutkuTopographic reorganization is an adaptive neural mechanism in which cortical neurons deprived of visual input begin to respond to stimuli from neighboring retinal locations. While traditionally studied in the context of permanent retinal lesions, recent research has increasingly utilized non-invasive, reversible, and short-term sensory-deprivation paradigms, such as monocular blindfolding, to examine this phenomenon in the adult primary visual cortex (V1). By employing a location-specific perceptual learning paradigm and fMRI, this study investigates the neural correlates of the transfer of perceptual learning to stimuli presented at an adjacent, untrained retinal location during sensory deprivation. By blindfolding the left eye, the cortical representation of the right eye’s blind spot in V1 was deprived of its typical visual input. Participants completed 15 to 20 days of training on a 2-interval forced-choice (2-IFC) orientation discrimination task presented within the blind spot (BS). Preliminary behavioral results from 16 participants showed significant improvements in orientation discrimination thresholds at the trained BS location, with a more pronounced transfer of learning to an adjacent location in the deprived condition than in the undeprived condition. fMRI sessions were conducted with 10 participants before and after the perceptual training (i.e., pre-test and post-test) to measure BOLD responses within the BS region of interest (ROI) to stimuli presented at the BS and at an adjacent location during both deprived and undeprived conditions. Regarding the deprivation condition, our findings were in line with a previous report by Jamal and Dilks (2020), showing an increased BOLD signal in the BS ROI in response to neighboring stimuli in the deprived condition compared to the undeprived condition. However, this difference is only approached to the significance in the present study with a large effect size. This indicates that adjacent stimuli elicited at least numerically higher responses in the BS ROI in the deprived condition compared to the undeprived condition, suggesting a robust trend toward the recruitment of the deprived cortical region by neighboring retinal inputs. However, the further analysis did not indicate any differentiated BOLD response across trained and untrained orientations after the perceptual training. Thus, the fMRI results in the present study did not align with the behavioral transfer of the perceptual learning effect. These findings reveal that the adult V1 maintains a capacity for rapid topographic reorganization following sensory loss, though identifying the subtle neural signatures of orientation-specific learning transfer may require more sensitive and sophisticated analysis techniques.Item Embargo An automated approach towards code comment smell detection and repair(2026-02) Çağlar, Hatice KübraCode comments are essential for software comprehension and maintenance; however, low-quality or inconsistent comments, referred to as code comment smells, can mislead developers and reduce code quality. This thesis proposes a classification-based framework for the automatic detection and repair of inline code comment smells. An enhanced dataset of 2,211 labeled inline comments, extended with manually curated repairs, is used to evaluate four large language models (GPT-4omini, o3-mini, DeepSeek-V3, and Codestral-2501 ) under zero-shot and few-shot prompting strategies. Experimental results show that lightweight instructiontuned models achieve competitive detection performance, while code-specialized models demonstrate stronger repair capability. The findings highlight challenges related to dataset imbalance, prompt sensitivity, and error propagation between detection and repair stages. To demonstrate practical feasibility, the proposed framework is implemented as a prototype tool, SmellSolver, integrated into Azure DevOps pull request workflows. Although not yet evaluated through user studies, the prototype illustrates the industrial applicability of automated inline comment smell detection and repair.Item Open Access A framework for collaborative integration and effective querying of biological pathways in a graph database(2026-02) Muhammad, NoorBiological pathways are used to represent molecular interactions and cellular processes. They serve as an important medium for communicating and organizing biological knowledge in a structured manner. These pathway models are typically curated independently by researchers and focus on specific processes and disease contexts. As the biological knowledge continues to expand, researchers require methods to combine and analyze multiple pathway models effectively in order to obtain broader, system-level insights. However, integrating pathway data from different sources remains challenging due to differences in semantics, structural organization, and level of detail across independently developed models. While standard pathway representation formats like SBGN Process Description (SBGN-PD) provide a common visual and semantic language, these formats are designed for individual pathway modeling and visualization. They offer limited inherent support for incremental integration and performing expressive queries across integrated pathway models remains difficult. This thesis presents a graph-based unified pathway model designed to support the incremental integration of biological pathway data and enable graph traversal-based queries. The proposed model focuses on preserving essential semantics of SBGN-PD pathways while allowing pathway data to be incorporated piece by piece into a unified graph. A key aspect of the model is its support for matching the incoming pathway entities (nodes and edges) with the existing graph content, where the matching behavior is customizable through user-defined thresholds. Building on this model, the system supports traversal-based queries such as neighborhood exploration, common stream detection, and paths between entities. The model and its database design are realized using the Neo4j graph database and are integrated into the Newt Pathway editor, providing an end-to-end system for pathway integration, querying, and visualization. The system is evaluated as a case study on representative SBGN-PD (SBGNML) maps using contract-based correctness checks (e.g., referential integrity, duplicate/self-loop suppression, and traversal budget soundness) and lightweight performance measurements (runtime and returned subgraph size). We also report sensitivity to matching thresholds and characterize scaling behavior on maps ranging from small examples to a large stress case.Item Embargo Learning-based analysis of pull request–issue alignment with large language models(2026-02) Altunhan, Mustafa YasirAccurate alignment between pull requests (PRs) and corresponding issues is crucial for efficient software development and maintaining code quality, as misalignments can lead to reduced traceability, hindered defect localization, and decreased maintainability. This thesis aims to improve automated PR–issue alignment classification by leveraging fine-tuned large language models (LLMs) across multiple alignment categories, to investigate the effects of PR–issue fields on model predictions through interpretability analysis, and to demonstrate how LLM-based PR– issue alignment analysis can be integrated into a real-world code review workflow. The proposed methodology consists of dataset preparation, LLM fine-tuning, interpretability analysis, and system implementation. An existing dataset is extended and data augmentation techniques are applied to address class imbalance. Subsequently, GPT-4o is fine-tuned via instruction tuning, and several open-source LLMs—including CodeLlama-7B, CodeQwen1.5-7B, StableCode-3B, CodeGemma-7B, and DeepSeek-Coder-6.7B—are fine-tuned using classificationspecific model heads. In addition, interpretability analysis using Shapley Additive Explanations (SHAP) is conducted to examine the influence of PR–issue fields on the predictions of the best-performing open-source LLM. In addition to the modeling approach, this thesis presents the design and implementation of an LLM-based PR–issue alignment tool integrated into realworld software development workflows. The tool is implemented as an extension to Bitbucket and Jira: it automatically analyzes PR–issue pairs upon pull request updates, reports alignment predictions directly within pull request interfaces, and allows developers to override automated decisions with an explicit label and explanation. To support traceability and post-hoc analysis, the tool persistently stores model predictions, developer overrides, and the corresponding PR–issue artifacts (including the code diff) in a commit-scoped manner. Experimental results show that fine-tuned LLMs outperform baseline models, achieving average improvements of 6.16% in accuracy and 14.53% in F1-macro. CodeLlama-7B emerges as the best-performing fine-tuned LLM overall, demonstrating consistent performance across evaluation metrics. Interpretability analysis further reveals that code diffs, together with issue body and PR body contents, exert the greatest influence on alignment predictions. Overall, the findings demonstrate that fine-tuning substantially enhances PR– issue alignment classification while interpretability analysis provides actionable insights into the dataset features driving alignment decisions. Moreover, the implemented PR–issue alignment tool shows that LLM-based alignment analysis can be embedded into practical review workflows, supporting improved traceability, transparency, and decision-making in software engineering.Item Embargo Rhythms in the brain: oscillatory dynamics, functional neural networks and perception(2026-02) Akdoğan, İremPerception is not passively driven by sensory input but is an active process shaped by temporally organized neural dynamics that operate across multiple spatial and functional scales. It relies on the flexible allocation of limited neural resources to meet changing sensory and cognitive demands, as well as the dynamic coordination required to stabilize, prioritize, and integrate sensory information in complex environments. The current thesis advances the central claim that neural oscillations constitute a core mechanism for implementing this resource allocation, enabling the same underlying neural circuits to dynamically support perception across di!erent contexts, representational demands, and network scales. To address this framework, the present thesis employs electroencephalography (EEG) and integrates complementary analysis approaches that span local oscillatory dynamics, multivariate representational structure, and large-scale functional network organization. The first study (Chapter 2) demonstrates that perceived visibility under visual masking depends on the temporal stabilization of sensory representations rather than response magnitude alone. Increased post-stimulus phase coherence in the theta and alpha bands predicts conscious perception, indicating that oscillatory alignment governs the e!ective allocation of processing resources to weak or ambiguous sensory input. The second study (Chapter 3) extends this account by examining the e!ects of attentional demands in the visual field, showing that visibility and attention share a common neural code when resources are su”cient, but diverge under high attentional load. These findings reveal that oscillatory gain mechanisms operate within strict capacity limits, leading to adaptive reorganization of representational formats as task demands increase. The last study (Chapter 4) examines resource allocation at the systems level, demonstrating that successful audiovisual processing requires the dynamic recruitment of distributed cortical networks, mediated by long-range theta- and alpha-band connectivity that transiently integrates sensory and associative regions. Together, the findings support a unified view of perception as a dynamic, demand-driven process in which neural oscillations govern the allocation of limited processing resources across time, representational formats, and large-scale networks. By showing how the same frequency-specific mechanisms scale from local signal stabilization to system-level integration, this thesis provides a coherent framework for understanding how the brain flexibly balances segregation and integration to construct stable perceptual experience under varying sensory and cognitive demands.Item Open Access A framework for validation and error resolution of biological pathway maps(2026-02) Özgül, Yusuf ZiyaThe exponential growth of data in the contemporary age underscores the increas-ing importance of visual data analysis. Consequently, the effective visualization of large-scale datasets is a major requirement for enhancing their readability and comprehensibility. Relational data, which encompasses any structure repre-sentable by a set of nodes and edges, falls within the scope of graph visualization. This technique is applicable across numerous domains, including software engi-neering, network systems, and computational biology. In biology, pathways and interactions are frequently modeled as graphs. The Systems Biology Graphical Notation (SBGN) is a standardized language devel-oped by scientists to model and visualize complex biological systems using graph visualization principles. While several tools, such as Newt and SBGN-ed, exist for the visualization of SBGN pathways, a critical challenge lies in ensuring ad-herence to the rigorous set of validation rules mandated by the SBGN standard. SBGN-ed, for example, is capable of validating SBGN maps and detecting rule violations. This thesis presents the design and implementation of a novel framework, SyB-ValS. SyBValS is designed to first validate a given SBGN map to identify errors and then, critically, generate actionable suggestions for resolving each detected error. A key feature is the programmatic capability that allows a client to select and apply these suggestions, thereby transforming the erroneous map into an op-timally corrected state. SyBValS represents the first comprehensive framework to integrate both map validation and automated, selective error resolution. Selected portions of this thesis were refined with the assistance of Gemini to enhance linguistic clarity and overall readability. The use of this tool was strictly limited to grammatical and stylistic improvements; it was not employed to generate original technical content, experimental results, or research conclusions.Item Embargo Pairwise whole genome alignment using locally consistent parsing(2026-01) İlgün, EcemPairwise whole-genome alignment is a fundamental problem in computational biology, with applications in evolutionary analysis, variant discovery and comparative genomics. This work focuses on the massive scaling challenges in pangenome analysis by using a hierarchical sketching method based on Locally Consistent Parsing (LCP). On a scale of billions of base pairs, efficient alignment typically relies on the seed-chain-extend heuristic: find exact-matching sketches (seeds), chain them co-linearly, and extend into the gaps. Established tools use minimizers or maximal unique matches (MUMs); we instead use LCP cores, which offer complete coverage, consistent spacing, and fewer seeds at higher levels. Distributed and parallelized multiple genome alignment relies on efficiently partitioning the input genomes into smaller segments that can be processed independently. Existing partitioning methods often rely on maximal exact matches (MEMs), maximal unique matches (MUMs), or minimizers for sketching. However, for MEMs/MUMs, the alignment process is complicated by the O(m· log n) time required to find MEMs of size m in a string of size n. Similarly, minimizers exhibit drawbacks in their distribution patterns and frequencies due to their short length, leading to suboptimal partitioning in terms of computational and communication overhead. Compared to minimizers, Locally Consistent Parsing (LCP) can offer a more thorough and condensed representation of the input data by identifying “cores,” or brief genomic sequences that are consistently present across genomes. We develop a fast, parallelizable pairwise genome alignment framework that uses a hierarchical seed-chain-extend strategy: seed at one LCP level, chain and merge matches, find unaligned regions, and, for each region, recurse by seeding only that region at the next lower level until a minimum level is reached. LCP cores can be computed hierarchically in linear time, leading to more balanced computational loads. We integrated LCPtools with the ChainX-LCP chaining algorithm and evaluated on E. coli (K-12 vs Sakai) and human (GRCh38 vs CHM13); on the human genome our seeding completed in 68 h while Mumemto was still running after 540 h, demonstrating scalability for reference-grade assemblies.Item Open Access Controllable diffusion-based visual editing(2026-02) Ekin, YiğitAdvancements in generative networks have significantly improved visual generation, particularly for image and video editing applications. However, key challenges remain in achieving controllable editing. Diffusion inpainting models often hallucinate or re-insert the intended object during object removal, and text-tovideo diffusion models struggle to follow a desired motion pattern without sacrificing prompt alignment for motion conditioned generation. This thesis addresses these gaps through two interconnected studies. First, we introduce a backgroundfocused image conditioning framework for object removal that utilizes focused embeddings and proposes a suppression method for removing foreground concept in the conditioning signal. By explicitly using such conditioning, it prevents common failure modes such as foreground leakage and mask-shape-driven hallucinations. Second, we develop a motion-conditioned video generation and editing method that achieves successful motion transfer from a reference to the generated video. By directly updating the positional embeddings, it achieves high fidelity motion aligned generation without sacrificing the textual condition alignment. Together, these contributions advance controllable visual editing by demonstrating that pretrained generative models contain useful behaviors beyond their explicit training objectives, and that providing the right guidance can unlock robust control with improved fidelity, consistency, and user-directed precision.Item Open Access Privacy preserving split learning(2026-01) Shabbir, AqsaSplit Learning enables collaborative model training without sharing raw data; however, its traditional form remains vulnerable because plaintext intermediate activations and gradients can leak sensitive information. These leakages enable attacks such as input reconstruction, label and property inference, and model manipulation, undermining the privacy guarantees that split learning aims to provide. This thesis addresses these limitations by designing a privacy-preserving split learning system. The proposed design inverts the conventional workflow so that labels, loss computation, and backpropagation remain entirely on the client, while all server-side computation is performed in the encrypted domain using homomorphic encryption. As a result, the server never observes plaintext activations, labels, or gradients during training, eliminating known attack surfaces. To make encrypted split learning practical, the thesis introduces an estimator that models ciphertext noise growth, bootstrapping requirements, and end-to-end runtime as functions of network architecture and split placement. The estimator jointly captures encrypted server-side computation and plaintext client-side computation, enabling noise- and budget-aware split selection without exhaustive empirical profiling. Our contributions include: (i) identifying and analyzing the components of traditional split learning that lead to privacy leakage, (ii) designing an inverted split learning system that eliminates information leakage by executing all server-side computation over encrypted data, and (iii) developing an estimator that enables the efficient use of homomorphic encryption in split learning under cryptographic and computational constraints.Item Open Access Genome reconstruction in beacons using summary statistics(2026-01) Saleem, KousarGenomic data-sharing beacons, designed to safeguard individual privacy while promoting scientific discovery, remain critically vulnerable to sophisticated genome reconstruction attacks that leverage publicly released summary statistics. This thesis systematically advances the understanding and effectiveness of these attacks, challenging the assumption that releasing simple allele frequencies (AFs) is a secure protocol. The fundamental flaw lies in the beacon’s protocol to account for linkage disequilibrium (LD), which allows a malicious party to infer individual data from combined summary statistics. Our foundational contribution established the feasibility of this threat with a two-stage optimization-based algorithm that utilized public LD and AFs, achieving an F1-score of 70% and confirming the inherent privacy risk. Building upon this, the research introduces a more powerful methodology: a single-stage joint optimization framework that unifies the objectives of SNP correlation and allele frequency alignment. This formulation not only increases reconstruction performance to an average F1-score of 71.4% but also yields substantial computational savings: reconstructing 2,000 SNPs across 100 individuals now requires 7.4 hours instead of 10 hours, representing a 26% reduction in runtime. Collectively, these results provide compelling evidence of the increasing practicality and sophistication of genome reconstruction attacks against beacon protocols, underscoring the urgent need for the development of robust, adaptive, and correlation-aware defense mechanisms to protect the integrity and privacy of genomic data infrastructure.Item Open Access Design of non-monetary incentives for efficiency in selfish routing via strategic intersection control(2025-12) Saltan, YusufUrban transportation networks routinely suffer from inefficiencies caused by selfish routing, whereby individual drivers select routes that minimize their own travel time rather than overall system delay. This decentralized behavior leads to user equilibria that can significantly deviate from system-optimal flows. Although monetary tolls can theoretically eliminate such inefficiencies, their practical, political, and equity-related limitations motivate the development of alternative, non-monetary control mechanisms. This thesis develops and analyzes two intersection-based incentive mechanisms that leverage modern Autonomous Intersection Management (AIM) to influence route choices and steer selfish routing toward socially efficient outcomes without monetary transfers. The first mechanism, termed Strategic Priority-Based Scheduling (SPBS), introduces small, route-dependent priority adjustments at intersections, thereby inducing controlled, path-dependent waiting times. Analytical examples, including Pigou’s network, show that even minimal priority asymmetries can substantially reduce inefficiency. These insights are further validated through high-fidelity microscopic simulations, demonstrating the mechanism’s feasibility under realistic driving and queueing dynamics. The second mechanism generalizes this approach through an analytical framework based on timestamp offsets. Intersections apply small additive adjustments to vehicles’ effective arrival times, inducing path-dependent node delays while preserving uniqueness of equilibrium travel costs, even when multiple equilibrium flows exist. This structure enables a bilevel optimization formulation in which a system planner designs timestamp offsets while anticipating user-equilibrium responses. Calibration using simulation-generated intersection delay data for the Sioux Falls network yields realistic quartic node cost models, and large-scale numerical experiments show that timestamp-based incentives can eliminate up to 68% of the inefficiency at user equilibrium, even under tight operational constraints. Taken together, these results demonstrate that intersections, traditionally viewed as network bottlenecks, can be transformed into powerful non-monetary control instruments. By exploiting the capabilities of modern AIM, the proposed mechanisms provide practical, scalable, and analytically grounded tools for improving network-wide efficiency without relying on tolls or major infrastructure modifications.Item Open Access Single-entry raffles with cryptographic verifiability and privacy(2026-01) Bayramoğlu, KeremLotteries are an integral part of generating revenue for public initiatives through regulated selection mechanisms. However, traditional raffle systems often face challenges related to privacy, fairness, and verifiability. To address these challenges, this thesis presents a novel system architecture for conducting single-entry raffles that ensures privacy and fairness through verifiability. In this work, two distinct architectures are proposed: (i) Centralized Architecture: Participants receive random numbers and unique IDs, which are added to an RSA accumulator along with their rank. The final accumulator is published in a trusted space, allowing participants to verify their entries. A verifiable random number generator selects the winner, with inclusion proofs available via queries. (ii) Blockchain- Based Architecture: Certificate authority hashes are mapped to smart contract numbers, enabling verifiable winner selection and participant inclusion checks for a private, auditable raffle. Proposed system leverages RSA accumulators and verifiable random functions to maintain both transparency and confidentiality, and the winner selection process is both transparent and auditable, maintaining the integrity of the raffle. By offering both centralized and blockchain-based solutions, this approach provides flexibility while maintaining the core principles of fairness and privacy. The proposed raffle system guarantees fairness, verifiability, and privacy in a single entry setup without requiring a trusted third party, thereby establishing a secure and transparent approach to online raffle management. The implementation is available at https://github.com/ASAP-Bilkent/ private-decentralized-lottery.Item Open Access Nextstereo: directionally driven channel expansion gives adaptive real-time stereo(2026-01) Ekinci, Ekin BerkWe present NeXtStereo, a lightweight stereo disparity estimation network designed for real-time depth perception. NeXtStereo builds on Widened ConvNeXtV2 blocks that strengthen cost aggregation while leveraging the scalability and generalization behavior of the ConvNeXt family. In addition, we introduce Directionally Modulated Attention (DMA), a novel attention mechanism that incorporates geometric priors to modulate features using directional cues. Together, these components improve structural detail recovery in challenging regions such as object boundaries, thin structures, and texture-weak areas, without relying on heavy 3D aggregation stacks. We evaluate NeXtStereo on SceneFlow, KITTI 2012/2015, and Middlebury, where it achieves a favorable accuracy/efficiency trade-off among real-time models and improves cross-domain robustness, with NeXtStereo-L achieving the lowest > 2px error among the compared methods. We also study adaptation to the MS2 outdoor driving dataset and observe reliable transfer under fine-tuning. Furthermore, NeXtStereo demonstrates strong compatibility with convolutional Low-Rank Adaptation (LoRA), enabling parameterefficient domain adaptation with improved stability compared to relevant realtime stereo matching baselines. Finally, we analyze selective 3D cost aggregation via a targeted ablation that replaces the first 1/4-scale aggregation block with a 3D ConvNeXt-style cost aggregation operator, characterizing the resulting accuracy/ efficiency trade-offs.Item Open Access A reinforcement learning-based approach for dynamic privacy protection in genomic data sharing beacons(2026-01) Aghdam, Masoud PoorghaffarThe rise of genomic sequencing has led to significant privacy concerns due to the sensitive and identifiable nature of genomic data. The Beacon Project, initiated by the Global Alliance for Genomics and Health (GA4GH), was designed to enable privacy-preserving sharing of genomic information via an online querying system. However, studies have revealed that the protocol is vulnerable to membership inference attacks, which can expose the presence of individuals in sensitive datasets. Existing countermeasures often degrade system utility or fail to adapt to evolving attack strategies due to their static nature. To address this, we model the interaction between the beacon and the adversary as a Stackelberg game. In this formulation, the attacker acts as the leader who selects a query strategy to maximize inference, while the defender acts as the follower who optimizes the response honesty to minimize privacy loss while maintaining utility. However, classical game-theoretic solutions are computationally intractable due to the vast search space of genomic queries. In this study, we bridge this gap by presenting a dynamic learning-based framework to approximate these equilibrium strategies. We employ a multi-agent reinforcement learning environment to solve this continuous game, training an adaptive defense policy that regulates response honesty against a sophisticated adversary capable of strategic query ordering and behavioral mimicry. Unlike conventional static defenses, this mechanism is capable of adapting in real time, dynamically differentiating between legitimate and adversarial query patterns to apply tailored policies. Consequently, this method enhances both privacy and utility, effectively countering sophisticated and evolving threats.Item Embargo Hardware acceleration for adaptive gamma correction in embedded systems(2026-01) Sarıçam, İlaydaEnhancing images in low-light conditions is a critical task in various domains, including photography, security systems, military, and autonomous driving models. These fields often require image processing and analysis tasks in low-light images due to a lack of lighting sources and shadows. However, there are limitations and bottlenecks in low-light image enhancement, such as under-enhancement, over-enhancement, and high power consumption. This thesis introduces a region-wise adaptive gamma correction (AGC) method, which is a non-learning-based approach, to enhance the visibility of lowlight images. In this study, to select the optimal gamma value adaptively, the image is partitioned into regions based on detected edges and ridges. Then, the optimal gamma value for each region is computed from average intensity, brightness, luminance, and RGB values. As a result, the gamma correction is applied to each region separately. With this region-wise approach, under-enhancement and over-enhancement of the input image are prevented. Furthermore, our approach is tailored for low-light image enhancement tasks in power-limited systems. Therefore, our implementation uses low-power devices rather than high-performance GPUs and CPUs, as typically used in the literature. To evaluate our results and output images, we use Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), Mean Squared Error (MSE), Naturalness Image Quality Assessor (NIQE), runtime, and power consumption as evaluation metrics. Also, to observe the effectiveness of our approach and compare it with prior studies, we conducted experiments on two datasets, namely, LOL and MITAdobe FiveK. When compared with previous non-learning-based methods, our approach achieves a twofold improvement in PSNR. Furthermore, we reduce the power consumption of the low-light image enhancement by more than 250X.Item Embargo Digital microfluidics for biomedical applications(2026-01) Güngen, Murat AlpPoint-of-care (PoC) diagnostic technologies aim to reduce global healthcare disparities by enabling decentralized testing without reliance on advanced laboratory infrastructure. Despite significant progress, many PoC systems remain largely confined to academic research settings, limiting their clinical and societal impact. Digital microfluidics (DMF), a programmable microfluidic approach based on electrowetting-on-dielectric (EWOD), enables the two dimensional controlled manipulation of discrete droplets and offers substantial advantages in flexibility, reconfigurability, and functional integration over conventional continuousflow microfluidic platforms. These characteristics make DMF a promising technological foundation for PoC diagnostics. In the first part of this thesis, the capabilities of a commercially available DMF platform, OpenDrop, are explored for biomedical applications relevant to PoC testing. The platform is employed to perform extracellular vesicle isolation, enzyme-linked immunosorbent assays. In addition, OpenDrop is used to rapidly generate image-based datasets to evaluate the feasibility of applying an in-house, U-Net-based, computer vision framework for droplet detection and classification on DMF devices. Building upon these demonstrations, the second part of this thesis focuses on the design, characterization, and fabrication of a custom DMF platform, designated “Markut.” This development includes computational analysis of the Young–Lippmann equation to guide EWOD optimization, systematic electrowetting experiments conducted in both air and oil to assess dielectric material performance, and the realization of a functional device architecture informed by these results. To support molecular diagnostic applications, a temperature control module is integrated to enable loop-mediated isothermal amplification (LAMP) assays. Furthermore, computer vision–based colorimetric analysis and electrical impedance measurements are incorporated to reliably distinguish between positive and negative LAMP outcomes. Overall, this thesis demonstrates the feasibility and versatility of both commercially available and custom-built DMF platforms for PoC-relevant biomedical applications. The presented results highlight DMF as a robust and scalable technology with strong potential to facilitate the translation of microfluidic diagnostics from laboratory research toward practical, real-world deployment.Item Open Access Robust deep learning under distribution shift: invariant feature learning and reliable test-time adaptation(2026-01) Karimi, SaeidDeep learning models often suffer significant performance degradation when deployed in environments whose data distributions differ from those encountered during training. This distribution shift remains a central challenge for robust visual recognition. Although Domain Generalization (DG) strives to learn models that generalize to unseen domains without accessing target data, recent studies show that many DG techniques yield limited improvements over empirical risk minimization due to reliance on spurious, domain-specific features. To address this issue, the first part of this thesis introduces Specific Domain Training (SDT), a method that disentangles spurious and invariant features via specific-domain sampling, masking, and variance-aware weight averaging. SDT improves both theoretical robustness and practical performance on DG benchmarks. The second part of the thesis focuses on Test-Time Adaptation (TTA), which adapts a pretrained model to incoming test samples without labels. Existing TTA methods often rely on noisy pseudo-labels and fail to leverage informative structure from source domains. To mitigate these limitations, we develop three complementary approaches. SATA uses source-domain style statistics to identify style-invariant test samples, ensuring stable entropy minimization while regularizing unreliable samples through consistency constraints. AdaPAC leverages subclass prototypes extracted using class-specific clustering to capture intra-class structure, selecting test samples that align well with source clusters and adapting the model with prototype-guided contrastive objectives. Shift-ACT introduces shift-aware, classspecific dynamic thresholding based on confidence discrepancies between source and target distributions, enabling reliable sample selection under class-wise distribution shifts. Together, these contributions advance the reliability of DG and TTA by reducing reliance on spurious cues, improving sample selection, and enabling robust adaptation under distribution shifts.