Conceptos Básicos
The author proposes the ϵ-Mesh Attack as a novel method to manipulate points on their original mesh surfaces subtly while preserving structural integrity, focusing on realistic applications and surface preservation.
Resumen
The content introduces the ϵ-Mesh Attack, a unique adversarial point cloud attack method for facial expression recognition. It emphasizes preserving surface structure and subtlety in manipulations, contrasting with traditional aggressive attacks. Experimental results show promising performance in protecting surface integrity while reducing model accuracy.
Key points:
- Introduction of the ϵ-Mesh Attack for 3D facial expression recognition.
- Focus on preserving surface structure and subtlety in manipulations.
- Comparison with traditional aggressive attacks like PGD and PGD-L2.
- Experimental results showing protection of surface structure and reduced model accuracy.
Estadísticas
Our experiments show that the suggested two attack methods cost almost the same in terms of time, compared to other gradient-based attacks like PGD.
For L2 metric, our suggested perpendicular and central ϵ-mesh attacks have a distance of 0.71 and 0.63 respectively, while PGD and PGD-L2 attacks have 1.28 and 0.97.
For Chamfer distance, results are as following: 71.53 for perpendicular, 53.36 for central, 212.22 for PGD, 120.21 for PGD-L2.
Citas
"The novelty of our approach lies in its ability to manipulate the points on their original mesh surfaces subtly while preserving its structural integrity."
"Our emphasis was on preserving the underlying mesh structure in the given point cloud."