New "sponge attacks" exploit SNN energy efficiency for increased power consumption

Researchers have identified a new security vulnerability in Spiking Neural Networks (SNNs) that exploits their energy efficiency. Dubbed "sponge attacks," these methods can significantly increase the energy consumption of SNNs during inference by inflating spike activity. A per-sample attack can boost SynOps by 1.5-2.6x while maintaining prediction accuracy, and a universal attack can increase energy usage by 1.09-1.24x, posing a realistic threat to battery-powered edge systems. AI

IMPACT Highlights a novel security vulnerability in energy-efficient AI hardware, potentially impacting the design and deployment of edge AI systems.

RANK_REASON Academic paper detailing a new attack vector on SNNs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• IBM DVS Gesture
• Loihi-1
• NMNIST
• SNNs
• Spiking Neural Networks
• sponge attacks

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →