Abstract
dc:descriptionDeep learning systems have become ubiquitous across society, driving state-of-the-art performance in countless applications. However, such systems inevitably reflect the statistical characteristics and imperfections present in the training data. This can introduce various forms of bias into the learned representations and decision-making processes, influencing model behavior in ways that are not always aligned with the intended task. In particular, high co-occurrence between certain data attributes and target labels that does not reflect true causal relationships, namely spurious correlations, can lead to models that fail catastrophically when deployed in out-of-distribution environments where such correlations do not exist, and raise fairness issues when spurious attributes correspond to protected characteristics such as race, gender, or age. This dissertation investigates such types of biases, with a primary focus on computer vision, while also examining generalization across modalities and application-specific challenges in facial analysis. It addresses several key limitations of existing approaches, including their dependence on annotations related to the attributes introducing the bias, limited ability to handle multiple biases, fragmented evaluation practices, and uncertain transferability beyond natural image benchmarks. To this end, the thesis introduces a set of novel methodological contributions. First, it proposes FLAC, a fairness-aware representation learning framework that suppresses bias attribute and target class associations without requiring bias attribute annotations. Second, it presents BAdd, a simple yet effective approach that mitigates single and multi-attribute biases by injecting bias-capturing features during training to encourage bias-neutral representations. Third, it introduces MAVias, an open-set bias mitigation method that leverages foundation models to automatically discover and mitigate unknown visual biases. Beyond algorithmic advances, this dissertation presents VB-Mitigator, an open-source framework that integrates datasets, metrics, and mitigation methods, enabling reproducible and fair comparisons. Building on this, a large-scale cross-domain benchmark study evaluates bias mitigation methods across text, audio, medical imaging, and video, providing the first systematic evidence of their cross-domain and modality generalization capabilities. Finally, the thesis investigates fairness in facial image analysis, revealing intersectional demographic biases in face recognition and verification systems, proposing a framework for training fairer face verification models using synthetic data, and introducing FaceX, an explainability method that provides face region-level and patch-level insights into face attribute classifiers, supporting bias diagnosis. Overall, this dissertation advances the state-of-the-art in bias mitigation by combining theoretical, methodological, empirical, and application-driven contributions. It offers practical tools, standardized benchmarks, and new insights that collectively contribute to the development of fairer, more robust, and more transparent artificial intelligence systems.
Degree
thesis:*- Grantor dc:publisher
- Harokopio University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Sarridis, Ioannis
- Σαρρίδης, Ιωάννης
Subjects
dc:subject × 16- Δίκαιη μηχανική μάθηση
- Βαθιά μάθηση
- Υπολογιστική όραση
- Μεροληψία
- Ψευδείς συσχετίσεις
- Fairness-aware machine learning
- Deep learning
- Computer vision
- Bias
- Spurious Correlations
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/61411