Unadversarial Examples: Designing Objects for Robust Vision

Authors: Hadi Salman, Andrew Ilyas, Logan Engstrom, Sai Vemprala, Aleksander Madry, Ashish Kapoor

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our framework yields improved performance on standard benchmarks, a simulated robotics environment, and physical-world experiments.
Researcher Affiliation Collaboration Hadi Salman hady@mit.edu MIT Andrew Ilyas ailyas@mit.edu MIT Logan Engstrom engstrom@mit.edu MIT Sai Vemprala saihv@microsoft.com Microsoft Research Aleksander M adry madry@mit.edu MIT Ashish Kapoor akapoor@microsoft.com Microsoft Research
Pseudocode Yes Unadversarial patches (cf. Algorithm 1 in Appendix A). Unadversarial textures (cf. Algorithm 2 in Appendix A).
Open Source Code Yes 1Our code is available at https://github.com/microsoft/unadversarial. The code link is in the abstract and Appendix C.6.
Open Datasets Yes Using the algorithm described in Section 2.3, we construct unadversarial patches of varying size for pre-trained Res Net-50 classifiers on the CIFAR [Kri09] and Image Net [RDS+15] datasets.
Dataset Splits Yes For evaluation, we add these patches at random positions, scales, and orientations to validation set images (see Appendix C for the exact protocol). We collect meshes corresponding to four Image Net classes: warplane, minibus, container ship, and trailer truck, from sketchfab.com. We generate a texture for each object using the unadversarial texture algorithm of Section 2.3, using the Image Net validation set as the set of backgrounds for the algorithm, and a pre-trained Res Net-50 as the classifier.
Hardware Specification Yes Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See Appendix C.5.
Software Dependencies No The paper mentions using 'Mitsuba' as a 3D renderer but does not provide specific version numbers for this or any other software dependencies.
Experiment Setup Yes For a more detailed account of each experimental setup, see Appendix C. Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See Appendix C.