# Researchers report attacks that defeat two on-device model protections

_Published Monday, October 19, 2026 at 10:10 AM EDT · Science, Security · Latest · Tier 2 — Notable_

Researchers report in an arXiv preprint that attacks on GroupCover and ArrowCloak recovered near-complete model functionality, undermining protections intended to prevent theft of proprietary models deployed on devices. Attack accuracy exceeded 2.4 times the baseline measured with full shielding and black-box access.

The researchers identify a weakness in transformations that obscure only the output side of model weights: the secret used for protection is uniquely determined. Their proposed BiShield-TEE framework obscures both input and output dimensions, leaving multiple equally valid candidate keys.

In their experiments, BiShield-TEE held attack accuracy to the full-shielding baseline and delivered higher inference efficiency than prior methods on real hardware.

## Sources

- [cs.CR updates on arXiv.org](https://arxiv.org/abs/2610.21565)

---
Canonical: https://techandbusiness.org/newswire/D510pgcr6j9q0easWYfSCY
Published: 2026-10-19T14:10:29.674Z
Story chronology: 2026-10-19T04:00:00.000Z
Retrieved: 2026-10-19T18:27:45.981Z
Publisher: Tech & Business (techandbusiness.org)