Lipschitz-Certifiable Training with a Tight Outer Bound
Authors: Sungyoon Lee, Jaewook Lee, Saerom Park
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In the experiments, we show that BCP achieves a tighter outer bound than the global Lipschitz-based outer bound. |
| Researcher Affiliation | Academia | Sungyoon Lee Seoul National University Seoul, Korea goman1934@snu.ac.kr Jaewook Lee Seoul National University Seoul, Korea jaewook@snu.ac.kr Saerom Park Sungshin Women s University Seoul, Korea psr6275@sungshin.ac.kr |
| Pseudocode | Yes | Algorithm 1 Box Constraint Propagation (BCP) Certifiable Training |
| Open Source Code | Yes | Our code is available at https://github.com/sungyoon-lee/bcp. |
| Open Datasets | Yes | We demonstrate that the proposed method can provide a tight outer bound for ℓ2-perturbation set and train certifiably robust networks, comparing its performance against state-of-the-art certifiable training methods (LMT [36], CAP [40], and IBP [13]) on MNIST and CIFAR10. Moreover, we also show that the BCP scheme can scale to Tiny Image Net and obtain a meaningful verification accuracy. |
| Dataset Splits | No | The paper mentions using MNIST, CIFAR-10, and Tiny Image Net datasets but does not explicitly provide the train/validation/test split percentages or counts within the main text. |
| Hardware Specification | Yes | We evaluate the computation times on a single Titan X GPU. |
| Software Dependencies | No | The paper does not specify software dependencies with version numbers (e.g., specific libraries, frameworks, or solvers with their versions). |
| Experiment Setup | Yes | Similar to Kurakin et al. [19], we train on a mixture of normal logit z(x) and the worst logit z(x) as follows: J (f, D; λ) = E(x,y) D (1 λ)L(z(x), y) + λL(z(x), y) i . We gradually increase the perturbation ϵ from 0 to the target bound ϵtarget and increase λ in (13) from 0 to 1, stabilizing the initial phase of training and improving natural accuracy [13, 44]. We use the same batch size for both methods as 50 on MNIST and CIFAR-10 and 5 on Tiny Image Net. In the case of WRN [42] on CIFAR-10, we can speed up to 61.1 sec/epoch using batch size of 128. |