Search for: All records

Snow and ice topography impact and are impacted by fluxes of mass, energy, and momentum in Arctic sea ice. We measured the topography on approximately a 0.5 km²drifting parcel of Arctic sea ice on 42 separate days from 18 October 2019 to 9 May 2020 via Terrestrial Laser Scanning (TLS). These data are aligned into an ice-fixed, lagrangian reference frame such that topographic changes (e.g., snow accumulation) can be observed for time periods of up to six months. Usingin-situmeasurements, we have validated the vertical accuracy of the alignment to ± 0.011 m. This data collection and processing workflow is the culmination of several prior measurement campaigns and may be generally applied for repeat TLS measurements on drifting sea ice. We present a description of the data, a software package written to process and align these data, and the philosophy of the data processing. These data can be used to investigate snow accumulation and redistribution, ice dynamics, surface roughness, and they can provide valuable context for co-located measurements.

 

Snow plays an essential role in the Arctic as the interface between the sea ice and the atmosphere. Optical properties, thermal conductivity and mass distribution are critical to understanding the complex Arctic sea ice system’s energy balance and mass distribution. By conducting measurements from October 2019 to September 2020 on the Multidisciplinary drifting Observatory for the Study of Arctic Climate (MOSAiC) expedition, we have produced a dataset capturing the year-long evolution of the physical properties of the snow and surface scattering layer, a highly porous surface layer on Arctic sea ice that evolves due to preferential melt at the ice grain boundaries. The dataset includes measurements of snow during MOSAiC. Measurements included profiles of depth, density, temperature, snow water equivalent, penetration resistance, stable water isotope, salinity and microcomputer tomography samples. Most snowpit sites were visited and measured weekly to capture the temporal evolution of the physical properties of snow. The compiled dataset includes 576 snowpits and describes snow conditions during the MOSAiC expedition.

 

Current machine learning models suffer from evasion attacks (i.e., adversarial examples) raising concerns in security-sensitive settings such as autonomous vehicles. While many countermeasures may look promising, only a few withstand rigorous evaluation. Recently, defenses using random transformations (RT) have shown impressive results, particularly BaRT (Raff et al. 2019) on ImageNet. However, this type of defense has not been rigorously evaluated, leaving its robustness properties poorly understood. The stochasticity of these models also makes evaluation more challenging and many proposed attacks on deterministic models inapplicable. First, we show that the BPDA attack (Athalye, Carlini, and Wagner 2018) used in BaRT’s evaluation is ineffective and likely overestimates its robustness. We then attempt to construct the strongest possible RT defense through the informed selection of transformations and Bayesian optimization for tuning their parameters. Furthermore, we create the strongest possible attack to evaluate our RT defense. Our new attack vastly outperforms the baseline, reducing the accuracy by 83% compared to the 19% reduction by the commonly used EoT attack (4.3× improvement). Our result indicates that the RT defense on Imagenette dataset (ten-class subset of ImageNet) is not robust against adversarial examples. Extending the study further, we use our new attack to adversarially train RT defense (called AdvRT). However, the attack is still not sufficiently strong, and thus, the AdvRT model is no more robust than its RT counterpart. In the process of formulating our defense and attack, we perform several ablation studies and uncover insights that we hope will broadly benefit scientific communities studying stochastic neural networks and their robustness properties. 

Intelligent voice assistants may soon become proactive, offering suggestions without being directly invoked. Such behavior increases privacy risks, since proactive operation requires continuous monitoring of conversations. To mitigate this problem, our study proposes and evaluates one potential privacy control, in which the assistant requests permission for the information it wishes to use immediately after hearing it. To find out how people would react to runtime permission requests, we recruited 23 pairs of participants to hold conversations while receiving ambient suggestions from a proactive assistant, which we simulated in real time using the Wizard of Oz technique. The interactive sessions featured different modes and designs of runtime permission requests and were followed by in-depth interviews about people's preferences and concerns. Most participants were excited about the devices despite their continuous listening, but wanted control over the assistant's actions and their own data. They generally prioritized an interruption-free experience above more fine-grained control over what the device would hear. 

Adversarial training (AT) has become a popular choice for training robust networks. However, it tends to sacrifice clean accuracy heavily in favor of robustness and suffers from a large generalization error. To address these concerns, we propose Smooth Adversarial Training (SAT), guided by our analysis on the eigenspectrum of the loss Hessian. We find that curriculum learning, a scheme that emphasizes on starting “easy” and gradually ramping up on the “difficulty” of training, smooths the adversarial loss landscape for a suitably chosen difficulty metric. We present a general formulation for curriculum learning in the adversarial setting and propose two difficulty metrics based on the maximal Hessian eigenvalue (H-SAT) and the softmax probability (P-SAT). We demonstrate that SAT stabilizes network training even for a large perturbation norm and allows the network to operate at a better clean accuracy versus robustness trade-off curve compared to AT. This leads to a significant improvement in both clean accuracy and robustness compared to AT, TRADES, and other baselines. To highlight a few results, our best model improves normal and robust accuracy by 6% and 1% on CIFAR-100 compared to AT, respectively. On Imagenette, a ten-class subset of ImageNet, our model outperforms AT by 23% and 3% on normal and robust accuracy respectively. 

Adversarial examples are a widely studied phenomenon in machine learning models. While most of the attention has been focused on neural networks, other practical models also suffer from this issue. In this work, we propose an algorithm for evaluating the adversarial robustness of k-nearest neighbor classification, i.e., finding a minimum-norm adversarial example. Diverging from previous proposals, we propose the first geometric approach by performing a search that expands outwards from a given input point. On a high level, the search radius expands to the nearby higher-order Voronoi cells until we find a cell that classifies differently from the input point. To scale the algorithm to a large k, we introduce approximation steps that find perturbation with smaller norm, compared to the baselines, in a variety of datasets. Furthermore, we analyze the structural properties of a dataset where our approach outperforms the competition. 

Abstract. Wind-driven redistribution of snow on sea ice alters itstopography and microstructure, yet the impact of these processes on radarsignatures is poorly understood. Here, we examine the effects of snowredistribution over Arctic sea ice on radar waveforms and backscattersignatures obtained from a surface-based, fully polarimetric Ka- and Ku-bandradar at incidence angles between 0∘ (nadir) and 50∘.Two wind events in November 2019 during the Multidisciplinary drifting Observatory forthe Study of Arctic Climate (MOSAiC) expedition are evaluated. During both events, changes in Ka- andKu-band radar waveforms and backscatter coefficients at nadir are observed,coincident with surface topography changes measured by a terrestrial laserscanner. At both frequencies, redistribution caused snow densification atthe surface and the uppermost layers, increasing the scattering at theair–snow interface at nadir and its prevalence as the dominant radar scattering surface. The waveform data also detected the presence of previousair–snow interfaces, buried beneath newly deposited snow. The additionalscattering from previous air–snow interfaces could therefore affect therange retrieved from Ka- and Ku-band satellite altimeters. With increasingincidence angles, the relative scattering contribution of the air–snowinterface decreases, and the snow–sea ice interface scattering increases.Relative to pre-wind event conditions, azimuthally averaged backscatter atnadir during the wind events increases by up to 8 dB (Ka-band) and 5 dB (Ku-band). Results show substantial backscatter variability within the scanarea at all incidence angles and polarizations, in response to increasingwind speed and changes in wind direction. Our results show that snowredistribution and wind compaction need to be accounted for to interpretairborne and satellite radar measurements of snow-covered sea ice.

 

Abstract. The remoteness and extreme conditions of the Arctic make it a very difficult environment to investigate. In these polar regions covered by sea ice, the wind is relatively strong due to the absence of obstructions and redistributes a large part of the deposited snow mass, which complicates estimates for precipitation hardly distinguishable from blowing or drifting snow. Moreover, the snow mass balance in the sea ice system is still poorly understood, notably due to the complex structure of its surface. Quantitatively assessing the snow distribution on sea ice and its connection to the sea ice surface features is an important step to remove the snow mass balance uncertainties (i.e., snow transport contribution) in the Arctic environment. In this work we introduce snowBedFoam 1.0., a physics-based snow transport model implemented in the open-source fluid dynamics software OpenFOAM.We combine the numerical simulations with terrestrial laser scan observations of surface dynamics to simulate snow deposition in a MOSAiC (Multidisciplinary Drifting Observatory for the Study of Arctic Climate) sea ice domain with a complicated structure typical for pressure ridges. The results demonstrate that a large fraction of snow accumulates in their vicinity, which compares favorably against scanner measurements. However, the approximations imposed by the numerical framework, together with potential measurement errors (precipitation), give rise to quantitative inaccuracies, which should be addressed in future work. The modeling of snow distribution on sea ice should help to better constrain precipitation estimates and more generally assess and predict snow and ice dynamics in the Arctic.