Phd Degree / Doktora

Permanent URI for this collectionhttps://hdl.handle.net/11147/2869

Browse

Filters

2017 - 201912020 - 20243

Settings

Subject: search.filters.subject.Machine learning

Search Results

Now showing 1 - 4 of 4
  • Doctoral Thesis
    Improving Low-Budget Semi-Supervised Approaches for Model Extraction Attacks
    (01. Izmir Institute of Technology, 2024) Genç, Didem; Baştanlar, Yalın; Tomur, Emrah
    Makine öğrenimi (ML) modelleri, etkinlikleri nedeniyle birçok alanda yaygın olarak kullanılmaktadır; ancak yüksek doğruluğa sahip modelleri eğitmenin maliyeti de yüksektik. Bu bağlamda, MLaaS (Machine Learning as a Service) platformları, API'ler aracılığıyla erişilebilen bulut tabanlı kara kutu modeller sunarak, model çalma saldırıları gibi güvenlik sorunlarını gündeme getirmektedir. Model çalma saldırıları, bulutta konuşlandırılmış bir makine öğrenimi modelini yalnızca kara kutu sorgulamalarıyla kopyalamayı amaçlamaktadır. Bu tez çalışmasında, etiketlenmemiş veriye erişimin kolay olduğu ancak etiketli verinin maliyetli olduğu senaryolarda, maliyet etkin ve yüksek doğruluklu bir model çalma saldırısı geliştirilmiştir. Literatürde sentetik veri setleri oluşturma, doğal veri setlerinden aktif öğrenme ile veri seçme ve yarı denetimli öğrenme gibi stratejiler önerilmektedir. Bu çalışmada ise, API üzerindeki kara kutu bir modele saldırmak için öz-denetimli öğrenen modellerden faydanılması önerilmiştir. Bu yöntemde, saldırganın geniş bir etiketlenmemiş veri havuzuna erişimi olduğu varsayılmakta ve bu veri, öz-denetimli SimCLR modelini eğitmek için kullanılmaktadır. Etiketsiz veri kümesinden belirli bir alt küme seçilir ve hedef modele sorgular gönderilerek bu veriler etiketlenir. Bu işlem sonucunda transfer veri seti oluşturulur. İlk ikame model, transfer veri setiyle SimCLR encoder'ına eklenen bir çok katmanlı algılayıcı (MLP)'nın ince ayar yapılarak eğitilmesi ile elde edilir. İkame modelin doğruluğunu artırmak için kalan etiketlenmemiş verilere otomatik etiketleme uygulanır; yüksek güvenli çıktılar doğrudan etiket olarak kullanılırken, düşük güvenli çıktılar hedef modelin etiketlediği örneklerle olan benzerliğe göre etiketlenir. Bu süreç, modelin karmaşık örüntüleri öğrenmesini ve veri çeşitliliğini artırmasını sağlayarak ikame modelin doğruluğunu hedef modele yaklaştıracak şekilde artırır. Önerilen methodun verimliliği CIFAR10 ve SVHN datasetleri üzerinde deneyler yapılarak verilmiştir.
  • Doctoral Thesis
    Development of Computational Models To Predict the Toxicity of Advanced Materials
    (01. Izmir Institute of Technology, 2023) Bilgi, Eyüp; Karakuş, Ceyda Öksel; Bedir, Erdal
    The aim of this study is to harness computational power to enhance existing knowledge on NM safety and to optimize the use of existing nanotoxicity data. The primary goal is to support the safe(r)-by-design concept, necessitating early integration of safety considerations into NM design through structural manipulation strategies. This thesis focuses on three case studies: zinc oxide, silver, and gold NP, using data manually collected from the literature. Analyses with zinc oxide and silver NP revealed a correlation between their toxicity and both internal (intrinsic properties, size, shape, surface charge) and external (cell and analysis-related properties) factors. For zinc oxide, it was found that coating had significant influence on cell viability, with a critical threshold identified at 20 µg/ml concentration and 10 nm size. Similarly, for silver NPs, concentration, size, and exposure time were significant factors. Coating with organic macromolecules increased cell viability, whereas green-synthesized NPs (using bacteria, plant extracts, algae) decreased it. The gold NP study highlighted that ensemble methods were more effective in elucidating complex relationships, with cellular uptake linked to particle size, zeta potential, concentration, and exposure time. Overall, this thesis contributes to safer-by-design strategies, crucial for developing commercially viable and safe NMs. The findings advocate for a broader toxicity evaluation approach, considering various physicochemical aspects and experimental procedures. The complex interactions observed suggest that advanced algorithms are necessary for accurate modeling, supporting the optimization of experimental parameters in NP engineering for biomedical applications.
  • Doctoral Thesis
    Planar Geometry Estimation With Deep Learning
    (Izmir Institute of Technology, 2022) Uzyıldırım, Furkan Eren; Özuysal, Mustafa
    Understanding the geometric structure of any scene is one of the oldest problems in Computer Vision. Most scenes include planar regions that provide information about the geometric structure and their automatic detection and segmentation plays an important role in many computer vision applications. In recent years, convolutional neural network architectures have been introduced for piece-wise planar segmentation. They outperform the traditional approaches that generate plane candidates with 3D segmentation methods from the explicitly reconstructed 3D point cloud. However, most of the convolutional neural network architectures are not designed and trained for outdoor scenes, because they require manual annotation, which is a time-consuming task that results in a lack of training data. In this thesis,we propose and develop a deep learning based framework for piece-wise plane detection and segmentation of outdoor scenes without requiring manually annotated training data. We exploit a network trained on imagery with annotated targets and an automatically reconstructed point cloud from either Structure from Motion-Multi View Stereo pipeline or monocular depth estimation network to estimate the training ground truth on the outdoor images in an iterative energy minimization framework. We show that the resulting ground truth estimate of various sets of images in the outdoor domain is good enough to improve network weights of different architectures trained on ground truth annotated images. Moreover, we demonstrate that this transfer learning scheme can be repeated multiple times iteratively to further improve the accuracy of plane detection and segmentation on monocular images of outdoor scenes.
  • Doctoral Thesis
    Computational Establishment of Microrna Metabolic Networks
    (Izmir Institute of Technology, 2017) Saçar Demirci, Müşerref Duygu; Allmer, Jens
    MicroRNAs (miRNAs) are single-stranded, small, non-coding RNAs, that control gene expression at the post transcriptional level through various mechanisms such as translational inhibition, degradation and destabilisation of their target mRNAs. Despite the fact that thousands of miRNAs have been reported in various species, most still remain unknown. Due to this, the identification of new miRNAs is an essential process for analysing miRNA mediated post transcriptional regulation mechanisms. Moreover, many biological approaches suffer from limitations in their capacity to reveal rare miRNAs, and are further restricted to the state of the organism under examination. Such limitations have resulted in the construction of sophisticated computational tools for identification of possible miRNAs in silico. However, these programs suffer from low sensitivity and/or accuracy and as a result they do not provide enough confidence for validating all their predictions experimentally. In this study, the aim is overcoming these challenges by creating a new and adaptable machine learning based method to predict potential miRNAs in any given sequence. The efficiency of proposed method is shown by comparison with available tools on various data sets. By using this approach, miRNAs from the genomes of various organisms like human (Homo sapiens), fly (Drosophila melanogaster) and tomato (Solanum lycopersicum) are identified. Moreover, networks between the possible miRNAs of virus and human genes as well as the communications among nuclear and organelle genomes of Solanum lycopersicum through miRNAs are investigated.