Neural networks are increasingly deployed on edge devices such as mobile phones and IoT hardware, where they offer low latency and better privacy than cloud-based alternatives. But this convenience comes with a security trade-off: the models themselves are physically present on the device, making them accessible to extraction attacks.
The paper's central warning is that such exposure does not stay contained. If an attacker successfully recovers a model from one device, that knowledge could be leveraged against other devices running the same or similar models. A single compromise could thus ripple across an entire ecosystem of edge hardware.
The authors argue that current security thinking focuses too narrowly on individual devices. Instead, they propose rethinking edge neural network security across devices, treating model lineage and shared architecture as part of the threat surface. This suggests a need for defenses that account for cross-device information leakage, not just local tamper resistance.