Adversarial Robustness with Partial Isometry.

Despite their remarkable performance, deep learning models still lack robustness guarantees, particularly in the presence of adversarial examples. This significant vulnerability raises concerns about their trustworthiness and hinders their deployment in critical domains that require certified levels of robustness. In this paper, we introduce an information geometric framework to establish precise robustness criteria for l2 white-box attacks in a multi-class classification setting. We endow the output space with the Fisher information metric and derive criteria on the input-output Jacobian to ensure robustness. We show that model robustness can be achieved by constraining the model to be partially isometric around the training points. We evaluate our approach using MNIST and CIFAR-10 datasets against adversarial attacks, revealing its substantial improvements over defensive distillation and Jacobian regularization for medium-sized perturbations and its superior robustness performance to adversarial training for large perturbations, all while maintaining the desired accuracy.

Metaheuristic for Optimal Dynamic K-Coloring Application on Band Sharing for Automotive Radars.

The number of vehicles equipped with radars on the road has been increasing for years and is expected to reach 50% of cars by 2030. This rapid rise in radars will likely increase the risk of harmful interference, especially since radar specifications from standardization bodies (e.g., ETSI) provide requirements in terms of maximum transmit power but do no mandate specific radar waveform parameters nor channel access scheme policies. Techniques for interference mitigation are thus becoming very important to ensure the long-term correct operation of radars and upper-layer ADAS systems that depend on them in this complex environment. In our previous work, we have shown that organizing the radar band into time-frequency resources that do not interfere with each other vastly reduces the amount of interference by facilitating band sharing. In this paper, a metaheuristic is presented to find the optimal resource sharing between radars, knowing their relative positions and thereby the line-of-sight and non-line-of-sight interference risks during a realistic scenario. The metaheuristic aims at optimally minimizing interference while minimizing the number of resource changes that radars have to make. It is a centralized approach where everything about the system is known (e.g., the past and future positions of the vehicles). This and the high computational load induce that this algorithm is not meant to be used in real-time. However, the metaheuristic approach can be extremely useful for finding near optimal solutions in simulations, allowing for the extraction of efficient patterns, or as data generation for machine learning.

Flight safety during Covid-19: A study of Charles de Gaulle airport atypical energy approaches.

During the COVID-19 period and particularly during lockdown, deviations from nominal operations have shown to become more frequent. To confirm this observation this paper proposes to evaluate the impact of COVID-19, and more generally of crises that lead to a sharp drop in traffic, on the pilot/controller system, especially during the critical approach and landing phases. To study the influence of this type of crisis on flight operations at Charles De Gaulle airport, an existing energy atypicality metric is applied on a reference period before COVID-19 and compared to the COVID-19 period. Whereas the traffic at Charles De Gaulle airport has decreased by around 90% on April 2020, the obtained statistics underlined an increase in the atypical flight ratio of around 50%. This trend can be explained in part by the appearance of glide interceptions from above as a result of trajectory shortenings, and an increase in the proportion of high speed approaches.

Impact of Covid-19 on passengers and airlines from passenger measurements: Managing customer satisfaction while putting the US Air Transportation System to sleep.

The COVID-19 pandemic has had a significant impact on the air transportation system worldwide. This paper aims at analyzing the effect of the travel restriction measures implemented during the COVID-19 pandemic from a passenger perspective on the US air transportation system. Four metrics based on data generated by passengers and airlines on social media are proposed to measure how the travel restriction measures impacted the relation between passengers and airlines in close to real-time. The proposed metrics indicate that each airline has reacted differently to the COVID-19 travel restriction measures from a passenger perspective, therefore they can be used by airlines and passengers to improve their decision making process. This report comes ahead of official data related to the same sequence of events, thereby showing the value of passenger-borne data in an industry where corporate priorities, institutional prudence, and passenger satisfaction come close together.