Field nodes that run image classifiers locally expose a valuable asset—the trained model—to theft, according to a new paper on arXiv. The abstract opens by stating this risk directly, framing the model as a first-class target for attackers who gain physical access to the device.

The paper notes that common defenses, including trusted-execution enclaves, collaborative-inference splits, and encrypted storage, all assume the device owner is the adversary from the outset. That assumption may not hold in scenarios involving physical capture, where the attacker controls the hardware and can probe it directly.

The title, "When Rotation Bounds Nothing," hints that certain rotation-based protections offer no security guarantee in this setting. The abstract does not yet reveal the proposed solution or experimental results, but it clearly positions physical capture as a blind spot in existing model-protection strategies.