Indiscriminate Data Poisoning Attacks Against Supervised Learning
Lightning Talk
In this talk I will summarise some of the work done in my PhD over the past few years on data poisoning. Our focus is on indiscriminate data poisoning attacks in worst-case scenarios against supervised learning, considering the machine learning pipeline: data sanitisation, hyperparameter learning, and training. We propose a novel attack formulation that considers the effect of the attack on the model’s hyperparameters. We apply this attack formulation to several ML classifiers using L2 regularisation. Our evaluation shows the benefits of using regularisation to help mitigate poisoning attacks, when hyperparameters are learnt using a trusted dataset. We then propose a novel stealthy attack formulation against regression models via multiobjective bilevel optimisation, where the two objectives are attack effectiveness and detectability. We experimentally show that state-of-the-art defences do not mitigate these stealthy attacks.