Stability Quantification of Neural Networks. (Quantification de la stabilité des réseaux de neurones)

Kavya Gupta

Abstract: Artificial neural networks are at the core of recent advances in Artificial Intelligence. One of the main challenges faced today, especially by companies likeThales designing advanced industrial systems is to ensure the safety of newgenerations of products using these technologies. In 2013 in a key observation, neural networks were shown to be sensitive to adversarial perturbations, raising serious concerns about their applicability in critically safe environments. In the last years, publications studying the various aspects of this robustness of neural networks, and rising questions such as "Why adversarial attacks occur?", "How can we make the neural network more robust to adversarial noise?", "How to generate stronger attacks?" etc., have grown exponentially. The contributions of this thesis aim to tackle such problems. The adversarial machine learning community concentrates majorly on classification scenarios, whereas studies on regression tasks are scarce. Our contributions bridge this significant gap between adversarial machine learning and regression applications.The first contribution in Chapter 3 proposes a white-box attackers designed to attack regression models. The presented adversarial attacker is derived from the algebraic properties of the Jacobian of the network. We show that our attacker successfully fools the neural network and measure its effectiveness in reducing the estimation performance. We present our results on various open-source and real industrial tabular datasets. Our analysis relies on the quantification of the fooling error as well as different error metrics. Another noteworthy feature of our attacker is that it allows us to optimally attack a subset of inputs, which may help to analyze the sensitivity of some specific inputs. We also, show the effect of this attacker on spectrally normalised trained models which are known to be more robust in handling attacks.The second contribution of this thesis (Chapter 4) presents a multivariate Lipschitz constant analysis of neural networks. The Lipschitz constant is widely used in the literature to study the internal properties of neural networks. But most works do a single parametric analysis, which do not allow to quantify the effect of individual inputs on the output. We propose a multivariate Lipschitz constant-based stability analysis of fully connected neural networks allowing us to capture the influence of each input or group of inputs on the neural network stability. Our approach relies on a suitable re-normalization of the input space, intending to perform a more precise analysis than the one provided by a global Lipschitz constant. We display the results of this analysis by a new representation designed for machine learning practitioners and safety engineers termed as a Lipschitz star. We perform experiments on various open-access tabular datasets and an actual Thales Air Mobility industrial application subject to certification requirements.The use of spectral normalization in designing a stability control loop is discussed in Chapter 5. A critical part of the optimal model is to behave according to specified performance and stability targets while in operation. But imposing tight Lipschitz constant constraints while training the models usually leads to a reduction of their accuracy. Hence, we design an algorithm to train "stable-by-design" neural network models using our spectral normalization approach, which optimizes the model by taking into account both performance and stability targets. We focus on Small Unmanned Aerial Vehicles (UAVs). More specifically, we present a novel application of neural networks to detect in real-time elevon positioning faults to allow the remote pilot to take necessary actions to ensure safety.