NEW YORK — What happened
Academic security researchers have published a vulnerability brief detailing power-analysis side-channel attacks capable of extracting plaintext inputs from on-device Neural Processing Units (NPUs) in consumer hardware.
Why it matters
As edge AI becomes standard on smartphones and laptops, users assume local processing guarantees privacy. This research proves that physical hardware leaks information even when data never touches an external cloud server.
Deep dive
By measuring minute voltage fluctuations across the silicon substrate during matrix multiplication operations, attackers can reconstruct character sequences being processed by local language models. The attack requires physical proximity or a compromised low-privilege background app monitoring power sensor APIs.
Report check (claims vs what is verified vs still rumor)
Researchers claim all current-generation consumer silicon is vulnerable. Verified: The attack works successfully on three popular mobile system-on-chips. Rumor: Chipmakers claim a simple firmware update can completely close the side-channel vector without performance loss.
Open questions
How will hardware architects redesign future neural chips to incorporate constant-power execution paths without destroying energy efficiency?