BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Sponge Examples: Energy-Latency Attacks on Neural Networks - Ilia 
 Shumailov\, University of Cambridge
DTSTART:20201020T130000Z
DTEND:20201020T140000Z
UID:TALK153277@talks.cam.ac.uk
CONTACT:Jack Hughes
DESCRIPTION:The high energy costs of neural network training and inference
  led to the use of acceleration hardware such as GPUs and TPUs. While this
  enabled us to train large-scale neural networks in datacenters and deploy
  them on edge devices\, the focus so far is on average-case performance. I
 n this work\, we introduce a novel threat vector against neural networks w
 hose energy consumption or decision latency are critical. We show how adve
 rsaries can exploit carefully crafted \, which are inputs designed to maxi
 mise energy consumption and latency.\n\nWe mount two variants of this atta
 ck on established vision and language models\, increasing energy consumpti
 on by a factor of 10 to 200. Our attacks can also be used to delay decisio
 ns where a network has critical real-time performance\, such as in percept
 ion for autonomous vehicles. We demonstrate the portability of our malicio
 us inputs across CPUs and a variety of hardware accelerator chips includin
 g GPUs\, and an ASIC simulator. We conclude by proposing a defense strateg
 y which mitigates our attack by shifting the analysis of energy consumptio
 n in hardware from an average-case to a worst-case perspective.\n\nhttps:/
 /arxiv.org/abs/2006.03463\n\nRECORDING : Please note\, this event will be 
 recorded and will be available after the event for an indeterminate period
  under a CC BY -NC-ND license. Audience members should bear this in mind b
 efore joining the webinar or asking questions.
LOCATION:Webinar
END:VEVENT
END:VCALENDAR