Tight Semidefinite Relaxations for Verifying Robustness of Neural Networks
Fuente:
arXiv
Enregistré dans:
| Auteurs principaux: | , , |
|---|---|
| Format: | Preprint |
| Publié: |
2025
|
| Sujets: | |
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866916688458416128 |
|---|---|
| author | Azuma, Godai Kim, Sunyoung Yamashita, Makoto |
| author_facet | Azuma, Godai Kim, Sunyoung Yamashita, Makoto |
| contents | For verifying the safety of neural networks (NNs), Fazlyab et al. (2019) introduced a semidefinite programming (SDP) approach called DeepSDP. This formulation can be viewed as the dual of the SDP relaxation for a problem formulated as a quadratically constrained quadratic program (QCQP). While SDP relaxations of QCQPs generally provide approximate solutions with some gaps, this work focuses on tight SDP relaxations that provide exact solutions to the QCQP for single-layer NNs. Specifically, we analyze tightness conditions in three cases: (i) NNs with a single neuron, (ii) single-layer NNs with an ellipsoidal input set, and (iii) single-layer NNs with a rectangular input set. For NNs with a single neuron, we propose a condition that ensures the SDP admits a rank-1 solution to DeepSDP by transforming the QCQP into an equivalent two-stage problem leads to a solution collinear with a predetermined vector. For single-layer NNs with an ellipsoidal input set, the collinearity of solutions is proved via the Karush-Kuhn-Tucker condition in the two-stage problem. In case of single-layer NNs with a rectangular input set, we demonstrate that the tightness of DeepSDP can be reduced to the single-neuron NNs, case (i), if the weight matrix is a diagonal matrix. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2504_09934 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Tight Semidefinite Relaxations for Verifying Robustness of Neural Networks Azuma, Godai Kim, Sunyoung Yamashita, Makoto Optimization and Control 62M45, 90C20, 90C22, 90C25, 90C26 For verifying the safety of neural networks (NNs), Fazlyab et al. (2019) introduced a semidefinite programming (SDP) approach called DeepSDP. This formulation can be viewed as the dual of the SDP relaxation for a problem formulated as a quadratically constrained quadratic program (QCQP). While SDP relaxations of QCQPs generally provide approximate solutions with some gaps, this work focuses on tight SDP relaxations that provide exact solutions to the QCQP for single-layer NNs. Specifically, we analyze tightness conditions in three cases: (i) NNs with a single neuron, (ii) single-layer NNs with an ellipsoidal input set, and (iii) single-layer NNs with a rectangular input set. For NNs with a single neuron, we propose a condition that ensures the SDP admits a rank-1 solution to DeepSDP by transforming the QCQP into an equivalent two-stage problem leads to a solution collinear with a predetermined vector. For single-layer NNs with an ellipsoidal input set, the collinearity of solutions is proved via the Karush-Kuhn-Tucker condition in the two-stage problem. In case of single-layer NNs with a rectangular input set, we demonstrate that the tightness of DeepSDP can be reduced to the single-neuron NNs, case (i), if the weight matrix is a diagonal matrix. |
| title | Tight Semidefinite Relaxations for Verifying Robustness of Neural Networks |
| topic | Optimization and Control 62M45, 90C20, 90C22, 90C25, 90C26 |
| url | https://arxiv.org/abs/2504.09934 |