Improving Neural Network Robustness via Persistency of Excitation

Sridhar, Kaustubh; Sokolsky, Oleg; Lee, Insup; Weimer, James

doi:10.23919/acc53348.2022.9867880

Cited by 5 publications

(3 citation statements)

References 14 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…1. in [72] shows that a model having a higher robust accuracy than other models at one ε level does not imply that such model is also more robust at other levels. The clean-robust accuracy trade-off [73,74] may also be interpreted similarly 19 -they are just most robust to different ε levels.…”

Section: Robust Accuracy Is Not a Good Robustness Metricmentioning

confidence: 92%

Optimization and Optimizers for Adversarial Robustness

Liang¹,

Buyun²,

Peng³

et al. 2023

Preprint

View full text Add to dashboard Cite

Empirical robustness evaluation (RE) of deep learning models against adversarial perturbations entails solving nontrivial constrained optimization problems. Existing numerical algorithms that are commonly used to solve them in practice predominantly rely on projected gradient, and mostly handle perturbations modeled by the 1, 2 and ∞ distances. In this paper, we introduce a novel algorithmic framework that blends a general-purpose constrained-optimization solver PyGRANSO with Constraint-Folding (PWCF), which can add more reliability and generality to the state-of-the-art RE packages, e.g., AutoAttack. Regarding reliability, PWCF provides solutions with stationarity measures and feasibility tests to assess the solution quality. For generality, PWCF can handle perturbation models that are typically inaccessible to the existing projected gradient methods; the main requirement is the distance metric to be almost everywhere differentiable. Taking advantage of PWCF and other existing numerical algorithms, we further explore the distinct patterns in the solutions found for solving these optimization problems using various combinations of losses, perturbation models, and optimization algorithms. We then discuss the implications of these patterns on the current robustness evaluation and adversarial training.

show abstract

Section: Robust Accuracy Is Not a Good Robustness Metricmentioning

confidence: 92%

Optimization and Optimizers for Adversarial Robustness

Liang¹,

Buyun²,

Peng³

et al. 2023

Preprint

View full text Add to dashboard Cite

show abstract

“…When adversaries utilize identical models of the targeted NN to attack it, it is considered a black box attack. As a result of those attacks, various strategies for increasing neural network robustness have been proposed, including adaptive gradient control [30], stability training [31], and adversarial learning [32], in which the NN is retrained to withstand these attacks.…”

Section: Robustness Verification Of Neural Networkmentioning

confidence: 99%

Security Robustness Evolution of Split Edge-Cloud Neural Networks

DOUCH,

ABID,

ZINE-DINE

et al. 2023

Preprint

View full text Add to dashboard Cite

Cloud computing, with its novel presence at the cloud Edge, proves indispensable to the success of the fifth (5G) and sixth-generation (6G) networks. This due to the myriad set of expected applications and required stringent constraints on latency and bandwidth, e.g., eHealth and Automotive vehicles which fall under the 5G URLLC (Ultra Reliable and Low Latency Communication) set. In order to reduce bandwidth usage at the Cloud core network and allow for faster latency responses, edge computing proposes a new servicing paradigm consisting of bringing computation and intelligence near where the end users reside, i.e., at the edge of the cloud. This is achieved by the edge infrastructure’s ability to support a variety of intelligent and computational jobs at the radio access network (RAN) level. However, many AI tasks, particularly those involving deep learning, pose a challenge as they require significantly more memory and processing capacity, which cannot be provisioned at the edge. Only cloud data centers can do. To address this issue, different edge intelligence methods, such as quantization, pruning, distributed inference, etc., have been proposed in the literature. This research investigates split neural networks (SNN), a neural network architecture with numerous early exits. SNN greatly decreases memory and computing, making it a promising NN architecture for edge devices and applications. Moreover, as the use of these SNNs grows, it becomes increasingly essential to study and validate their safety properties, notably their tolerance to small perturbations, before deploying them in safety-critical applications. This paper presents, as far as we know, the first exploratory work on the robustness assessment of split Edge Cloud neural networks. We evaluated SNN robustness using auto_LiRPA an advance neural network verification tool that is base on bounds calculation. We also compared the relative robustness of SNN to a normal NN, considering the many parameters that influence SNN’s robustness. Our experiment results show that SNN not only reduces the average inference time by 3/4. But it also proved to be four to ten times more resilient against adversarial attacks than normal NN.

show abstract

“…The classes of systems where the PE property is used include, but are not restricted to, those appearing in problems of identification [10], adaptive control [11,12], model identification [13], learning-based identification [14], and state estimation [15,16]. For instance, the so-called gradient systems, which appear in the context of gradient-descent estimation algorithms are among the linear time-varying systems where PE is necessary and sufficient for UES of the origin.…”

Section: Introductionmentioning

confidence: 99%

Hybrid Persistency of Excitation in Adaptive Estimation for Hybrid Systems

Saoud¹,

Maghenem²,

Lorı́a³

et al. 2023

Preprint

View full text Add to dashboard Cite

We propose a framework of stability analysis for a class of linear non-autonomous hybrid systems, with solutions evolving in continuous time governed by an ordinary differential equation and undergoing instantaneous changes governed by a difference equation. Furthermore, the jumps may also be triggered by exogeneous hybrid signals. The proposed framework builds upon a generalization of notions of persistency of excitation (PE) and uniform observability (UO), which we redefine to fit the realm of hybrid systems. Most remarkably we propose for the first time in the literature a definition of hybrid persistency of excitation. Then, we establish conditions, under which, hybrid PE implies hybrid UO and, in turn, uniform exponential stability (UES) and input-to-state stability (ISS). Our proofs rely on an original statement for hybrid systems, expressed in terms of Lp bounds on the solutions. We also demonstrate the utility of our results on generic adaptive estimation problems. The first one concerns the so-called gradient systems, reminiscent of the popular gradient-descent algorithm. The second one pertains to the design of adaptive observers/identifiers for a class of hybrid systems that are nonlinear in the input and in the output, and linear in the unknown parameters. In both cases, we illustrate through meaningful examples that the proposed hybrid framework succeeds in scenarii where the classical purely continuousor discrete-time counterparts fail.

show abstract

Improving Neural Network Robustness via Persistency of Excitation

Cited by 5 publications

References 14 publications

Optimization and Optimizers for Adversarial Robustness

Optimization and Optimizers for Adversarial Robustness

Security Robustness Evolution of Split Edge-Cloud Neural Networks

Hybrid Persistency of Excitation in Adaptive Estimation for Hybrid Systems

Contact Info

Product

Resources

About