Expressive Losses for Verified Robustness via Convex Combinations

Authors: Alessandro De Palma, Rudy R Bunel, Krishnamurthy Dj Dvijotham, M. Pawan Kumar, Robert Stanforth, Alessio Lomuscio

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We present a comprehensive experimental evaluation of CC-IBP, MTL-IBP and Exp-IBP on benchmarks from the robust vision literature, demonstrating that the three methods attain state-of-the-art performance in spite of their conceptual simplicity. We now present experimental results that support the centrality of expressivity, as defined in 3, to verified training.
Researcher Affiliation Collaboration Alessandro De Palma1a, Rudy Bunel2, Krishnamurthy (Dj) Dvijotham2, M. Pawan Kumar2, Robert Stanforth2, Alessio Lomuscio3 1Inria, École Normale Supérieure, PSL University, CNRS 2Google Deep Mind 3Imperial College London
Pseudocode Yes Pseudo-code for both losses can be found in appendix D. Algorithm 1 provides pseudo-code for CC-MTL, MTL-IBP and Exp-IBP.
Open Source Code Yes Code is available at https://github.com/alessandrodepalma/expressive-losses.
Open Datasets Yes We benchmark on the CIFAR10 (Krizhevsky & Hinton, 2009), Tiny Image Net (Le & Yang, 2015), and downscaled Image Net (64 64) (Chrabaszcz et al., 2017) datasets. MNIST (Le Cun et al., 2010) is a 10-class (each class being a handwritten digit) classification dataset consisting of 28 28 greyscale images, with 60,000 training and 10,000 test points.
Dataset Splits Yes We train on the entire training sets, and evaluate MNIST and CIFAR-10 on their test sets, Image Net and Tiny Image Net on their validation sets.
Hardware Specification Yes In particular, we relied on the following CPU models: Intel i7-6850K, i9-7900X CPU, i9-10920X, i9-10940X, AMD 3970X. The employed GPU models for most of the experiments were: Nvidia Titan V, Titan XP, and Titan XP. The experiments of 6.3 and appendix G.5 on CIFAR-10 for ϵ = 8/255 were run on the following GPU models: RTX 4070 Ti, RTX 2080 Ti, RTX 3090.
Software Dependencies No The paper mentions several software components like PyTorch, automatic Li RPA, OVAL framework, and torchvision.datasets, but does not provide specific version numbers for them (e.g., 'We implemented CC-IBP, MTL-IBP and Exp-IBP in Py Torch').
Experiment Setup Yes Details pertaining to the employed datasets, hyper-parameters, network architectures, and computational setup are reported in appendix F. Table 3: Hyper-parameter configurations for the networks trained via CC-IBP, MTL-IBP and Exp-IBP from tables 1 and 4.