Search

Bit flipping attacks are one class of attacks on neural networks with numerous defense mechanisms invented to mitigate its potency. Due to the importance of ensuring the robustness of these defense mechanisms, we perform an empirical study on the Aegis framework. We evaluate the baseline mechanisms of Aegis on low-entropy data (MNIST), and we evaluate a pre-trained model with the mechanisms fine-tuned on MNIST. We also compare the use of data augmentation to the robustness training of Aegis, and how Aegis performs under other adversarial attacks, such as the generation of adversarial examples. We find that both the dynamic-exit strategy and robustness training of Aegis has some drawbacks. In particular, we see drops in accuracy when testing on perturbed data, and on adversarial examples, as compared to baselines. Moreover, we found that the dynamic exit-strategy loses its uniformity when tested on simpler datasets. The code for this project is available on GitHub., Comment: 9 pages, 6 figures, 3 tables

Simulating curvature due to gravity through warped surfaces is a common visualization aid in Physics education. We reprise a recent experiment exploring orbital trajectories on a precise 3D-printed surface to mimic Newtonian gravity, and elevate this analogy past the status of a mere visualization tool. We present a general analysis approach through which this straightforward experiment can be used to create a reasonably advanced computational orbital mechanics lab at the undergraduate level, creating a convenient hands-on, computational pathway to various non-trivial nuances in this discipline, such as the mean, eccentric, and true anomalies and their computation, Laplace-Runge-Lenz vector conservation, characterization of general orbits, and the extraction of orbital parameters. We show that while the motion of a marble on such a surface does not truly represent a orbital trajectory under Newtonian gravity in a strict theoretical sense, but through a proposed projection procedure, the experimentally recorded trajectories closely resemble the Kepler orbits and approximately respect the known conservation laws for orbital motion. The latter fact is demonstrated through multiple experimentally-recorded elliptical trajectories with wide-ranging eccentricities and semi-major axes. In this first part of this two-part sequence, we lay down the formal basis of this exposition, describing the experiment, its calibration, critical assessment of the results, and the computational procedures for the transformation of raw experimental data into a form useful for orbital analysis., Comment: Article 1 of a two-part sequence. Article 2 (applications) to be released shortly. Comments are most welcome !