Abstract
Talk prepared for the AISB Workshop on Explainability and Transparency in AI, XTAI 2022.
Machine Learning, in particular Deep Learning, has most recently provided the state-of-the-art results for many tasks such as object recognition, text-to-speech processing, and credit-card fraud detection. In many cases, Deep Learning has even out-performed human performance on these very same tasks. Despite this advance in performance, however, the existence of so-called adversarial examples is well known within the community. These adversarial examples are the metaphorical ‘blind-spot’ of Deep Learning models, where very small (often human-imperceptible) changes to model’s input can result catastrophic miss-classifications. These adversarial examples then pose a great safety risk, especially in systems where safety is critical such as fully-automotive vehicles.
To defend against and attempt to eradicate the existence of adversarial examples in Deep Learning models, principle works have sought to search for their existence within fixed-sized regions around training points, and use the found adversarial examples as a criterion for learning. These works have demonstrated how the robustness of Deep Learning models against adversarial examples improves through these training regimes.
Our work means to compliment and improve on these existing approaches by adapting the size of the searchable regions around training points, based upon the complexity of the problem and data sampling density. The result is each training point has an adapted region around it to which adversarial examples can be searched for and found.
We demonstrate how, through the development of uniquely-adaptive searchable regions, existing methods can help to further improve the robustness of Deep Learning models, and also make the existing methods applicable to non-image related tasks by providing an upper bound for discovering adversarial examples.
In this presentation, we will explore how adversarial examples can be determined through the use of existing approaches. Further to these approaches, how our method allows us to generate unique and adapted region sizes for all training points in a dataset.