Duration: 6 months, October 2019 - March 2020
Deadline: 26th July 2019
We request expressions of interest (one brief paragraph on the nature of your expertise in the area) for the ‘Physical Adversarial Camouflage’ project
This task would involve the development and demonstration of a physical adversarial attack, utilizing a dataset chosen in consultation with the Dstl technical partner that would be physically plausible to implement.
A key expectation of this task is the demonstration of the attack in a real, physical setting. One example of how this could be achieved are:
1. Using aerial image data of vehicles taken from a plane or drone or other source of live aerial imagery.
2. Using a scaled down table-top model vehicle (similar to an “airfix” kit, or 3D printed equivalent), with overhead camera at appropriate height for the scale model, and simulated atmospheric effects by e.g. controlling lighting, smoke, etc.
3. Run experiments to show the vehicle being correctly classified in multiple lighting conditions from multiple angles before modifications are applied, and incorrectly classified in multiple lighting conditions from multiple angles after modifications are applied.
The vehicle classifier setting is just one possibility; facial recognition is another; we are open to innovative suggestions of use cases in the proposals.
The results of this task should be outlined in a report that gives the details of experiments carried out and the technical details of the attack algorithm(s) used. The experiments performed should evaluate the attack under varied conditions (translations, rotations, lighting conditions, zooming in and out) and using a range of vehicles (or other targets). In addition to the report, working code and demonstrator software implementing the attack should also be provided.