arXiv subjects

Marcus Becker

Publications and source records attributed to Marcus Becker.

3 recordsLinked to original sources

arXiv2026

Adversarial Evasion Attacks on Computer Vision using SHAP Values

The paper introduces a white-box attack on computer vision models using SHAP values. It demonstrates how adversarial evasion attacks can compromise the performance of deep learning models by reducing output confidence or inducing misclassifications. Such attacks are particularly insidious as they can deceive the perception of an algorithm while eluding human perception due to their imperceptibility to the human eye. The proposed attack leverages SHAP values to quantify the significance of individual inputs to the output at the inference stage. A comparison is drawn between the SHAP attack and the well-known Fast Gradient Sign Method. We find evidence that SHAP attacks are more robust in generating misclassifications particularly in gradient hiding scenarios.

cs.CV↗

arXiv2025

Closed-loop model-predictive wind farm flow control under time-varying inflow using FLORIDyn

Wind farm flow control has been a key research focus in recent years, driven by the idea that a collectively operating wind farm can outperform individually controlled turbines. Control strategies are predominantly applied in an open-loop manner, where the current flow conditions are used to look up precomputed steady-state set points. Closed-loop control approaches, on the other hand, take measurements from the farm into account and optimize their set points online, which makes them more flexible and resilient. This paper introduces a closed-loop model-predictive wind farm controller using the dynamic engineering model FLORIDyn to maximize the energy generated by a ten-turbine wind farm. The framework consists of an Ensemble Kalman Filter to continuously correct the flow field estimate, as well as a novel optimization strategy. To this end the paper discusses two dynamic ways to maximize the farm energy and compares this to the current look-up table industry standard. The framework relies solely on turbine measurements without using a flow field preview. In a 3-hour case study with time-varying conditions, the derived controllers achieve an overall energy gain of 3 to 4.4 % with noise-free wind direction measurements. If disturbed and biased measurements are used, this performance decreases to 1.9 to 3 % over the greedy control baseline with the same measurements. The comparison to look-up table controllers shows that the closed-loop framework performance is more robust to disturbed measurements but can only match the performance in noise-free conditions.

eess.SY↗

arXiv2021

Applying and testing a novel method to estimate animal density from motion-triggered cameras

Estimating animal abundance and density are fundamental goals of many wildlife monitoring programs. Camera trapping has become an increasingly popular tool to achieve these monitoring goals due to recent advances in modeling approaches and the capacity to simultaneously collect data on multiple species. However, estimating the density of unmarked populations continues to be problematic due to the difficulty in implementing complex modeling approaches, low precision of estimates, and absence of rigor in testing of model assumptions and their influence on results. Here, we describe a novel approach that uses still image camera traps to estimate animal density without the need for individual identification, based on the Time spent In Front of the Camera (TIFC). Using results from a large-scale multi-species monitoring program with nearly 3,000 cameras deployed over six years in Alberta, Canada, we provide a reproducible methodology to estimate parameters and we test key assumptions of the TIFC model. We compare moose (Alces alces) density estimates from aerial surveys and TIFC, including incorporating correction factors for known TIFC assumption violations. The resulting corrected TIFC density estimates are comparable to aerial density estimates. We discuss the limitations of the TIFC method and areas needing further investigation, including the need for long-term monitoring of assumption violations and the number of cameras necessary to provide precise estimates. Despite the challenges of assumption violations and high measurement error, cameras and the TIFC method can provide useful alternative or complementary animal density estimates for multi-species monitoring when compared to traditional monitoring methods.

q-bio.QM↗