Year: 2025
Type: dissertation
Abstract: This thesis tackles the challenge of mitigating skin tone bias in deep learning models for skin cancer classification through the application and extension of Latent Adversarial Debiasing (LAD).While deep neural networks have demonstrated high performance in various tasks, they often adopt present biases inside the data unintentionally.To address this, we adapt LAD, an adversarial based technique for mitigating biases and evaluate its effectiveness across multiple architectures, including ResNet, DenseNet, and EfficientNet, under fully biased and semi-biased conditions.The study compares the models on a combination of the ISIC2020 and the PAD-UEFS, in addition to evaluating fairness metrics such as Equal Opportunity (EO), Disparate Impact (DI), and Accuracy Parity (AP) alongside traditional performance measures.Results show that LAD-enhanced models achieve significant improvements in fairness, with DenseNet121 and EfficientNetB0 in particular demonstrating robust performance across both fairness and accuracy.Moreover, the analysis highlights the trade-off between fairness and predictive accuracy, indicating that the model is sensitive to hyperparameters and extensive searches should be considered when considering implementing it.This work contributes insights into fairnessaware model design and provides a concrete foundation for deploying debiased deep learning models in meaningful real-world applications such as medical imaging.The implementation code is publicly available on the following github repository https
Authors: Βλαχαΐτης, Παναγιώτης Δ., Βλαχαΐτης, Παναγιώτης Δ.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7111561838