Interactions between Software Product Lines and Adversarial Machine - PowerPoint PPT Presentation

Interactions between Software Product Lines and Adversarial Machine Learning Paul TEMPLE 1 Gilles PERROUIN 1 , 2 Pierre-Yves SCHOBBENS 1 Patrick HEYMANS 1 1 NaDI, PReCISE, Faculty of Computer Science, University of Namur 2 FNRS April, 12 th 2019 this work is funded by the EOS VeriLearn project April, 12 th 2019 P TEMPLE (NaDI, PReCISE, UNamur) SPLs and advML 1 / 7

Software Product Lines and Machine Learning Machine Learning in Software Product Line ML can help to deal with # of products (Linux kernel ≈ 2 15 k ) Inference mechanism → no generation; reason on configurations April, 12 th 2019 P TEMPLE (NaDI, PReCISE, UNamur) SPLs and advML 2 / 7

Software Product Lines and Machine Learning Machine Learning in Software Product Line ML can help to deal with # of products (Linux kernel ≈ 2 15 k ) Inference mechanism → no generation; reason on configurations Performance prediction with ML Guo et al. , Variability-Aware performance prediction: a statistical learning approach , ASE 2013 Siegmund et al. , Performance-influence models for highly configurable systems , FSE 2015 Temple et al. , Learning-based performance specialization of configurable systems , SPLC 2016 April, 12 th 2019 P TEMPLE (NaDI, PReCISE, UNamur) SPLs and advML 2 / 7

Machine Learning Learning process Sample a (small) number of configurations Generate associated products and measure performances Build a prediction model to infer performances Use the model on new configurations April, 12 th 2019 P TEMPLE (NaDI, PReCISE, UNamur) SPLs and advML 3 / 7

Machine Learning Learning process Sample a (small) number of configurations Generate associated products and measure performances Build a prediction model to infer performances Use the model on new configurations Machine Learning assumptions Initial sample is representative of the configurations’ population New configurations follow the same distribution than initial sample April, 12 th 2019 P TEMPLE (NaDI, PReCISE, UNamur) SPLs and advML 3 / 7

Machine Learning But... April, 12 th 2019 P TEMPLE (NaDI, PReCISE, UNamur) SPLs and advML 4 / 7

Machine Learning But... April, 12 th 2019 P TEMPLE (NaDI, PReCISE, UNamur) SPLs and advML 4 / 7

Machine Learning But... Craft configurations to artificially increase the number of errors of the prediction model → Adversarial Machine Learning April, 12 th 2019 P TEMPLE (NaDI, PReCISE, UNamur) SPLs and advML 4 / 7

Breaking the rules of Machine Learning Adversarial Machine Learning Appeared in 2004 Popular around 2014 with GANs a Still popular today b Biggio and Roli, Wild patterns: Ten years after the rise of adversarial machine learning , Pattern Recognition,Vol. 84, 2018 a Goodfellow et al. , Generative Adversarial Nets , NIPS 2014 b Zhang et al. , DeepRoad: GAN-based Metamorphic Autonomous Driving System Testing , ASE’18 April, 12 th 2019 P TEMPLE (NaDI, PReCISE, UNamur) SPLs and advML 5 / 7

Breaking the rules of Machine Learning Adversarial Machine Learning Appeared in 2004 Popular around 2014 with GANs a Still popular today b Biggio and Roli, Wild patterns: Ten years after the rise of adversarial machine learning , Pattern Recognition,Vol. 84, 2018 a Goodfellow et al. , Generative Adversarial Nets , NIPS 2014 b Zhang et al. , DeepRoad: GAN-based Metamorphic Autonomous Driving System Testing , ASE’18 Goal of Adversarial Machine Learning Better understand ML algorithms, their assumptions and weaknesses April, 12 th 2019 P TEMPLE (NaDI, PReCISE, UNamur) SPLs and advML 5 / 7

Adversarial Machine Learning intuition Add perturbations ⇒ April, 12 th 2019 P TEMPLE (NaDI, PReCISE, UNamur) SPLs and advML 6 / 7

Adversarial Machine Learning intuition Add perturbations ⇒ April, 12 th 2019 P TEMPLE (NaDI, PReCISE, UNamur) SPLs and advML 6 / 7

Adversarial Machine Learning intuition Add perturbations ⇒ April, 12 th 2019 P TEMPLE (NaDI, PReCISE, UNamur) SPLs and advML 6 / 7

Adversarial Machine Learning intuition Add perturbations ⇒ April, 12 th 2019 P TEMPLE (NaDI, PReCISE, UNamur) SPLs and advML 6 / 7

SPL and Adversarial ML SPL for AdvML Various ways to create adversarial configurations → variability modeling Open questions: Is it interesting? Can those attacks be composed? Can it help designing new adversarial techniques? April, 12 th 2019 P TEMPLE (NaDI, PReCISE, UNamur) SPLs and advML 7 / 7

SPL and Adversarial ML SPL for AdvML Various ways to create adversarial configurations → variability modeling Open questions: Is it interesting? Can those attacks be composed? Can it help designing new adversarial techniques? AdvML for SPL testing (Submitted to SPLC’19) Use AdvML for SPL Quality Assurance Adapted one adversarial technique to SPL Understand if the Variability Model is under/over constrained How to take into account constraints from the SPL in the process? April, 12 th 2019 P TEMPLE (NaDI, PReCISE, UNamur) SPLs and advML 7 / 7