Edge Device Model Exposure Triggers Cross-Node Adversarial Transfer Attacks
Start here
On October 20, 2026, an arXiv paper re-examined the security problem of cross-device edge neural networks: exposure of a model on one device endangers other devices, and attackers can exploit adversarial example transferability between models to launch cross-node attacks. The researchers therefore propose the VeilNN defense framework, which applies covert and irreversible obfuscation at the network unit and parameter levels to reduce adversarial example transferability between on-device models. Under four types of adversarial attacks, VeilNN achieves an average defense success rate of 49.19%, up to 81.19%, an improvement of more than 33.16% over existing defenses, and is applicable to architectures such as MLP, CNN, and Vision Transformer.
Generated by AI from reports · Updated 23 days ago
Attention and developments
Attention, hourly
Now 22, 0% in 24 hours·Peak 22 (Nov 6 03:00)
Related research
This event's paper
- Model Exposure on One Device Endangers Others: Rethinking Edge Neural Network Security across Devices
arXiv · 2026-10-19 · Deep reading
Progress timeline
- arXiv: Adversarial Attacks, Backdoors & PoisoningExposure of a model on one device endangers other devices: revisiting cross-device edge neural network security
The researchers propose the VeilNN defense framework, which applies covert and irreversible obfuscation at the network unit and parameter levels to reduce the transferability of adversarial examples between edge-side models and defend against cross-node attacks that share the same model. Under four types of adversarial attacks, VeilNN achieves an average defense success rate of 49.19%, up to 81.19%, an improvement of more than 33.16% over existing defenses, and is applicable to architectures such as MLP, CNN, and Vision Transformer.