Isaac Dunn

Isaac Dunn, Hadrien Pouget, Daniel Kroening et al.

Existing methods for testing DNNs solve the oracle problem by constraining the raw features (e.g. image pixel values) to be within a small distance of a dataset example for which the desired DNN output is known. But this limits the kinds of faults these approaches are able to detect. In this paper, we introduce a novel DNN testing method that is able to find faults in DNNs that other methods cannot. The crux is that, by leveraging generative machine learning, we can generate fresh test inputs that vary in their high-level features (for images, these include object shape, location, texture, and colour). We demonstrate that our approach is capable of detecting deliberately injected faults as well as new faults in state-of-the-art DNNs, and that in both cases, existing methods are unable to find these faults.

Isaac Dunn, Laura Hanu, Hadrien Pouget et al.

We cannot guarantee that training datasets are representative of the distribution of inputs that will be encountered during deployment. So we must have confidence that our models do not over-rely on this assumption. To this end, we introduce a new method that identifies context-sensitive feature perturbations (e.g. shape, location, texture, colour) to the inputs of image classifiers. We produce these changes by performing small adjustments to the activation values of different layers of a trained generative neural network. Perturbing at layers earlier in the generator causes changes to coarser-grained features; perturbations further on cause finer-grained changes. Unsurprisingly, we find that state-of-the-art classifiers are not robust to any such changes. More surprisingly, when it comes to coarse-grained feature changes, we find that adversarial training against pixel-space perturbations is not just unhelpful: it is counterproductive.

Isaac Dunn, Hadrien Pouget, Tom Melham et al.

Neural networks are vulnerable to adversarially-constructed perturbations of their inputs. Most research so far has considered perturbations of a fixed magnitude under some $l_p$ norm. Although studying these attacks is valuable, there has been increasing interest in the construction of (and robustness to) unrestricted attacks, which are not constrained to a small and rather artificial subset of all possible adversarial inputs. We introduce a novel algorithm for generating such unrestricted adversarial inputs which, unlike prior work, is adaptive: it is able to tune its attacks to the classifier being targeted. It also offers a 400-2,000x speedup over the existing state of the art. We demonstrate our approach by generating unrestricted adversarial inputs that fool classifiers robust to perturbation-based attacks. We also show that, by virtue of being adaptive and unrestricted, our attack is able to defeat adversarial training against it.