A novel technique that boosts adversarial robustness by injecting perturbations from robust models into new tasks. Unlike traditional attacks, DDA enriches training with diverse, transferable adversarial signals, improving defenses across multiple datasets.

A benchmark for evaluating the robustness of monocular depth estimation models in endoscopic scenarios, featuring real-world image corruptions. Alongside it, we propose a new metric, mDERS, and release SCARED-C, a dataset tailored to assess performance under challenging endoscopic conditions.

A robust Federated Learning framework that leverages pre-trained models to improve kidney stone diagnosis while preserving data privacy. Our two-stage method—Learning Parameter Optimization and Federated Robustness Validation—boosts both accuracy and resilience to image corruptions across decentralized datasets.

A novel feature engineering method that combines meta-learning and causal analysis to automate the search for high-impact feature transformations. By encoding feature distributions and selecting causally relevant inputs, MACFE significantly improves prediction performance.

We present FedAgain, a federated learning algorithm for medical imaging anomaly detection across decentralized networks with heterogeneous data. The system uses a dual-signal trust mechanism based on reconstruction error and model divergence to downweight suspicious client contributions while maintaining privacy, achieving precision gains of up to +14.49% and F1 score improvements of up to +10.20% over traditional FedAvg.

A comparative evaluation between Vision Transformers and CNNs for kidney stone classification from endoscopic images. The ViT-base model pretrained on ImageNet-21k significantly outperformed ResNet50, achieving 95.2% accuracy on complex endoscopic patches compared to ResNet50's 64.5%, demonstrating that ViT architectures are a superior alternative for this medical imaging task.

This chapter explores the critical need for evaluating the robustness of deep learning models, especially in safety-sensitive applications. Focusing on adversarial attacks, we highlight the vulnerabilities of neural networks to small, targeted input changes and emphasize the importance of developing reliable techniques to ensure trustworthy AI systems.

A novel approach designed to improve generalization and manifold representation under distributional shifts. By combining a multi-branch master model with a student model in an offline distillation setup, our method achieves up to 40% performance gains in visual similarity tasks through more compact and robust clustering.