Many recent works have proposed methods to train classifiers with local robustness properties, which can provably eliminate classes of evasion attacks for most inputs, but not all inputs. Since data distribution shift is very common in security applications, e.g., often observed for malware detection, local robustness cannot guarantee that the property holds for unseen inputs at the time of deploying the classifier. Therefore, it is more desirable to enforce global robustness properties that hold for all inputs, which is strictly stronger than local robustness. In this paper, we present a framework and tools for training classifiers that satisfy global robustness properties. We define new notions of global robustness that are more suitable for security classifiers. We design a novel booster-fixer training framework to enforce global robustness properties. We structure our classifier as an ensemble of logic rules and design a new verifier to verify the properties. In our training algorithm, the booster increases the classifier's capacity, and the fixer enforces verified global robustness properties following counterexample guided inductive synthesis. We show that we can train classifiers to satisfy different global robustness properties for three security datasets, and even multiple properties at the same time, with modest impact on the classifier's performance. For example, we train a Twitter spam account classifier to satisfy five global robustness properties, with 5.4% decrease in true positive rate, and 0.1% increase in false positive rate, compared to a baseline XGBoost model that doesn't satisfy any property.

The uplift of the Tibetan Plateau is an important geological event, but there is considerable controversy about its growth history. Different geological observations contribute to this controversial issue, while data from geochemistry, tectonics, and paleontology further fuel the debate. Vertebrate fossils have provided significant evidence for documenting the uplift of the Tibetan Plateau in the geologic past. The earliest fossil evidence recently collected from the Oligocene Dingqing Formation in central Tibet includes the climbing perch and cyprinine fish fossils whose modern close relatives are distributed in the tropical zone of Asia and Africa. These discoveries not only are significant for the phylogeny and zoogeography of fishes, but also imply that the hinterland of the Tibetan Plateau was a warm and humid lowland at ~26 Ma. The co-existing plant assemblage, which includes palms and golden rain trees among others, indicates that the warm and humid airs from the Indian Ocean could flow deeply into central Tibet, consistent with the inference from the fish fossils. Since that time, the geographical features and natural environments within the Tibetan Plateau have greatly changed. The Tibetan Plateau was consistently uplifted in the Early Miocene and reached an elevation of ~3000 m, which was demonstrated by fish, mammal, and plant fossils. The endemic schizothoracines (snow carps) originated from the Miocene when the Tibetan Plateau turned into a barrier for mammalian migrations between north and south sides. A series of fish and mammal fossils provided unequivocal evidence that the Tibetan Plateau uplifted close to its modern elevation in the Pliocene and developed a cryospheric environment. As a result, the plateau region became the origination center for the cold-adapted Quaternary Ice Age fauna.

Previous research demonstrates that large basins on the periphery of the northern edge of the Tibetan Plateau were partitioned during development of intrabasin mountain ranges. These topographic barriers segregated basins with respect to surface flow and atmospheric circulation, ponded sediments, and formed rain shadows. However, complex mixing between airmasses and nonsystematic isotope-elevation lapse rates have hampered application of quantitative paleoaltimetry to determine the timing of development of critical topographic barriers. We address the timing and drivers for changes in surface connectivity and atmospheric circulation in the Linxia and Xunhua basins using a multidisciplinary approach incorporating detrital zircon geochronology, Monte Carlo inverse flexural modelling, and published stable isotope data. Disruption of surface flow between the two basins during exhumation of the Jishi Shan preceded development of topography sufficient to intercept moisture-bearing airmasses. Detrital zircon data point to disruption of an eastward-flowing axial fluvial network between 14.7 and 13.1 Ma, coincident with the onset of exhumation in the Jishi Shan. Flexural modelling suggests that by 13 Ma, the Jishi Shan had developed 0.3 ± 0.1 km of relief; sufficient to disrupt eastward-flowing drainage networks but insufficient to intercept moisture-bearing airmasses. Stable isotope data indicate that, although surface connections between the Xunhua and Linxia basins were broken, the two basins continued to be dominated by a common climate regime until 9.3 Ma. Subsequent reintegration of surface flow between the basins occurred between 9.3 and 7.6 Ma. Divergence in the stable isotope and depositional environment records between the two basins suggests that at 9.3 Ma the paleo-Yellow River breached the growing Jishi Shan dam, and may have reintegrated surface flow between the two basins via erosion of the modern Yellow River gorge, which transects the Jishi Shan. The reintegration of the Xunhua and Linxia basins’ surface connection is confirmed by reintroduction of a Songpan-Ganzi flysch sediment source by 7.6 Ma. Continued exhumation and uplift of the Jishi Shan developed 0.8 ± 0.2 km of relief by ca. 8 Ma capable of intercepting moisture-bearing airmasses; isolating and increasing aridity in the Xunhua Basin while decreasing it in the Linxia Basin. Our findings point to protracted development of the modern ca. 1 km of relief in the Jishi Shan between 14 and ca. 4.5 Ma followed by attainment of a topographic equilibrium which persists into modern times.

We name a rare, hypercarnivorous and durophagous mustelid Sonitictis moralesi, new genus and species, in honour of Jorge Morales for his contributions to carnivore palaeontology. Sonitictis moralesi is from the middle Miocene Tunggur Formation of Inner Mongolia, China (Tunggurian Land Mammal age). S. moralesi has a short and robust jaw that deepens slightly behind the m1 and possesses a shallow groove at its ventral surface. Dentally, it has robust premolars and highly hypercarnivorous m1 that has lost the metaconid and a trenchant talonid. Known by three jaw fragments, knowledge of this new mustelid is still poor and its relationship uncertain. However, based on its possession of a ventral groove, robust premolars and hypercarnivorous first molar, we tentatively place it in the subfamily Mellivorinae, which includes modern Mellivora capensis (honey badger) and extinct species of Eomellivora, plus others. We conducted mandibular force profile analyses on one S. moralesi specimen and compared bite force estimates to several specimens of E. piveteaui, E. wimani and M. capensis. The results suggest that Sonitictis likely was a more capable predator and consumer of vertebrate prey than the extant Mellivora, reaching the low end of the bite force estimates for the larger and more robust Eomellivora.

Photo-active shape memory polymers (SMPs) are considered as a promising candidate for converting light into mechanical energy. However, most known SMPs are only thermo-responsive. To achieve photo-activity, photo-responsive choromophores or fillers usually have to be incorporated from the very beginning of the material synthesis. Here, we introduce a novel post-synthesis approach to endow normal SMPs with photo-active properties using mussel-inspired surface chemistry. Without changing the original properties, the resultant polydopamine (PDA) coated SMPs show an efficient photo-active performance. The coating can be easily patterned and erased, which allows flexible light-triggered 3-D shape deformation of a planar SMP sheet. Moreover, owing to the high chemical activity, the PDA coating also provides a platform to optimize the surface properties of the photo-responsive SMPs through secondary surface modification.