Bridging the gap: developing robust and ethical AI systems

Abstract

Machine learning and deep learning algorithms have demonstrated exceptional performance across various computer vision tasks, including face detection, attribute prediction, object detection, image segmentation, and classification. The capabilities of deep neural networks (DNNs) have driven the development of numerous computer vision- based artificial intelligence (AI) systems, employed in diverse applications such as autonomous vehicles, COVID-19 detection in X-ray images, facial recognition for security and surveillance systems. Despite these advancements, the real-world applicability of such systems remains a critical challenge due to issues such as the unethical exploitation of AI technologies and failures of deep models on real-world datasets. One significant concern is the unethical exploitation of AI systems, particularly in contexts involving privacy violations and the generation of fake media. Instances such as unauthorized profiling for targeted advertising and the creation and dissemination of Deepfake content pose serious ethical and social risks. These activities not only infringe on individual privacy but also create widespread societal distress, potentially leading to psychological harm. Another challenge lies in the failure of deep models to generalize effectively to unseen real-world datasets. This is largely due to a lack of robustness and generalizability in model training, as well as data distribution shifts during testing. For example, a model trained on the Labeled Faces in the Wild (LFW) dataset containing over 13,000 unconstrained facial images collected from the web, may perform poorly when evaluated on the CelebA dataset, which consists of more than 200,000 celebrity face images annotated with 40 facial attributes. This performance drop occurs because of discrepancies in data characteristics. To address these challenges, this dissertation investigates the problems in the input space, the model space, and the feature space, identifying the input space as the root cause. The lack of a control mechanism to guide the attention of deep models toward meaningful image features underpins these issues. To resolve this, we propose novel solutions leveraging the dynamics of attention mechanisms in deep models. Two solutions are introduced to mitigate the unethical exploitation of AI systems: 1) Privacy Preservation: A framework that enables users to selectively anonymize facial identity and attribute information in images. This allows users to control the attention of the model, blocking it from processing specific attributes such as gender or age. 2) Deepfake Detection: An approach tailored to detect fake media across varying compression levels and algorithms used on social media platforms. This solution improves detection accuracy by training models to identify unseen artifacts in low-quality compressed media. The latter part of this dissertation addresses deep model failures on real-world datasets with two innovative solutions: 1) Data Shift Mitigation: A method to counteract data distribution shifts by learning uniform perturbations in the dataset. These perturbations improve model attention toward meaningful features, ensuring better performance in real-world scenarios. 2) Training Robustness: A strategy to diversify attention of the model during training, compelling it to learn overlooked features within the dataset. This approach enhances the robustness and generalizability of the model by expanding the range of features it considers. Overall, this dissertation presents novel mechanisms to address critical challenges in AI system reliability and ethical deployment by focusing on privacy preservation, Deep- fake detection, and generalizable learning. These solutions contribute significantly to the development of robust, generalizable, and ethical AI systems for real-world applications.