Adversarial Robustness in Machine Learning

Machine learning models trained on clean data routinely fail on inputs that have been intentionally perturbed to cause misprediction — a class of vulnerabilities known as adversarial examples. This challenge extends beyond deep neural networks to tree-based ensemble models and non-differentiable classifiers, and affects both inference-time attacks (evasion) and training-time attacks (data poisoning and multi-party learning corruption).

This project investigates adversarial robustness across model families, with particular attention to deep forests, gradient boosting, and other ensemble methods that resist simple gradient-based analysis. Research addresses both the characterisation of robustness under realistic threat models and the development of defences that remain effective when attackers have knowledge of the defence strategy. The work connects to broader concerns about the reliability of ML-based systems in safety-critical and adversarial deployment environments.