Liron David

Tsufit Shua, Liron David, Mahmood Sharif

Traffic-Sign Recognition (TSR) is a critical safety component for autonomous driving. Unfortunately, however, past work has highlighted the vulnerability of TSR models to physical-world attacks, through low-cost, easily deployable adversarial patches leading to misclassification. To mitigate these threats, most defenses focus on altering the training process or modifying the inference procedure. Still, while these approaches improve adversarial robustness, TSR remains susceptible to attacks attaining substantial success rates. To further the adversarial robustness of TSR, this work offers a novel approach that redefines traffic-sign designs to create signs that promote robustness while remaining interpretable to humans. Our framework takes three inputs: (1) A traffic-sign standard along with modifiable features and associated constraints; (2) A state-of-the-art adversarial training method; and (3) A function for efficiently synthesizing realistic traffic-sign images. Using these user-defined inputs, the framework emits an optimized traffic-sign standard such that traffic signs generated per this standard enable training TSR models with increased adversarial robustness. We evaluate the effectiveness of our framework via a concrete implementation, where we allow modifying the pictograms (i.e., symbols) and colors of traffic signs. The results show substantial improvements in robustness -- with gains of up to 16.33%--24.58% in robust accuracy over state-of-the-art methods -- while benign accuracy is even improved. Importantly, a user study also confirms that the redesigned traffic signs remain easily recognizable and to human observers. Overall, the results highlight that carefully redesigning traffic signs can significantly enhance TSR system robustness without compromising human interpretability.

Liron David, Avishai Wool

Human-chosen passwords are the a dominant form of authentication systems. Passwords strength estimators are used to help users avoid picking weak passwords by predicting how many attempts a password cracker would need until it finds a given password. In this paper we propose a novel password strength estimator, called PESrank, which accurately models the behavior of a powerful password cracker. PESrank calculates the rank of a given password in an optimal descending order of likelihood. PESrank estimates a given password's rank in fractions of a second---without actually enumerating the passwords---so it is practical for online use. It also has a training time that is drastically shorter than previous methods. Moreover, PESrank is efficiently tweakable to allow model personalization in fractions of a second, without the need to retrain the model; and it is explainable: it is able to provide information on why the password has its calculated rank, and gives the user insight on how to pick a better password. Our idea is to cast the question of password rank estimation in a probabilistic framework used in side-channel cryptanalysis. We view each password as a point in a $d$-dimensional search space, and learn the probability distribution of each dimension separately. The dimensions represent the base word, plus a dimension for each possible transformation such as adding a suffix or using a capitalization pattern. Using this model, password strength estimation is analogous to side-channel rank estimation. We implemented PERrank in Python and conducted an extensive evaluation study of it. We also integrated it into the registration page of a course at our university. Even with a model based on 905 million passwords, the response time was well under 1 second, with up to a 1-bit accuracy margin between the upper bound and the lower bound on the rank.