BMO: Boundary-aware model obfuscation defends against cross-node attacks
Start here
In October 2026, an arXiv paper proposes the boundary-aware model obfuscation method BMO to defend against cross-node attacks in edge device scenarios. The method is based on null-space constraints, decomposes model parameters into a boundary position part and a boundary curvature part, and then injects noise into the curvature component to generate differentiated decision boundaries. On four types of adversarial attacks and four edge-side models, BMO achieves an average defense success rate of 68.68%, while the strongest baseline reaches only 22.32%; benign accuracy is comparable to the baselines.
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
- Mind Your Boundaries: Boundary-Aware Model Obfuscation against Cross-node Attack
arXiv · 2026-10-19 · Deep reading
Progress timeline
- arXiv: Adversarial Attacks, Backdoors & PoisoningBMO: A Boundary-Aware Model Obfuscation Method for Cross-Node Attacks
BMO defends against cross-node attacks in edge-device scenarios by obfuscating model parameters and injecting noise into the curvature component of the decision boundary. The method decomposes model parameters into a boundary position part and a boundary curvature part based on a null-space constraint, then injects noise into the curvature component to generate differentiated decision boundaries. On four types of adversarial attacks and four edge-side models, BMO achieves an average defense success rate of 68.68%, while the strongest baseline reaches only 22.32%, with benign accuracy comparable to the baselines.