BEGIN:VCALENDAR VERSION:2.0 PRODID:-//Talks.cam//talks.cam.ac.uk// X-WR-CALNAME:Talks.cam BEGIN:VEVENT SUMMARY:Human-producible Adversarial Examples - David Khachaturov\, Univer sity of Cambridge DTSTART:20231128T140000Z DTEND:20231128T150000Z UID:TALK207199@talks.cam.ac.uk CONTACT:Hridoy Sankar Dutta DESCRIPTION:Visual adversarial examples have so far been restricted to pix el-level image manipulations in the digital world or have required sophist icated equipment such as 2D or 3D printers to be produced in the physical real world. We present the first ever method of generating human-producibl e adversarial examples for the real world that requires nothing more compl icated than a marker pen. We call them adversarial tags. First\, building on top of differential rendering\, we demonstrate that it is possible to b uild potent adversarial examples with just lines. We find that by drawing just 4 lines we can disrupt a YOLO-based model in 54.8% of cases\; increas ing this to 9 lines disrupts 81.8% of the cases tested. Next\, we devise a n improved method for line placement to be invariant to human drawing erro r. We evaluate our system thoroughly in both digital and analogue worlds a nd demonstrate that our tags can be applied by untrained humans. We demons trate the effectiveness of our method for producing real-world adversarial examples by conducting a user study where participants were asked to draw over printed images using digital equivalents as guides. We further evalu ate the effectiveness of both targeted and untargeted attacks\, and discus s various trade-offs and method limitations\, as well as the practical and ethical implications of our work.\n\nRECORDING : Please note\, this event will be recorded and will be available after the event for an indetermina te period under a CC BY -NC-ND license. Audience members should bear this in mind before joining the webinar or asking questions. LOCATION:Webinar &\; FW11\, Computer Laboratory\, William Gates Buildin g. END:VEVENT END:VCALENDAR