Timezone: »

Explanations can be manipulated and geometry is to blame
Ann-Kathrin Dombrowski · Maximillian Alber · Christopher Anders · Marcel Ackermann · Klaus-Robert Müller · Pan Kessel

Tue Dec 10 10:45 AM -- 12:45 PM (PST) @ East Exhibition Hall B + C #165

Explanation methods aim to make neural networks more trustworthy and interpretable. In this paper, we demonstrate a property of explanation methods which is disconcerting for both of these purposes. Namely, we show that explanations can be manipulated arbitrarily by applying visually hardly perceptible perturbations to the input that keep the network's output approximately constant. We establish theoretically that this phenomenon can be related to certain geometrical properties of neural networks. This allows us to derive an upper bound on the susceptibility of explanations to manipulations. Based on this result, we propose effective mechanisms to enhance the robustness of explanations.

Author Information

Ann-Kathrin Dombrowski (TU Berlin)
Maximillian Alber (TU Berlin)
Christopher Anders (Technische Universität Berlin)
Marcel Ackermann (HHI)
Klaus-Robert Müller (TU Berlin)
Pan Kessel (TU Berlin)

More from the Same Authors