skip to main content

Title: Interpretable Detection of Distribution Shifts in Learning Enabled Cyber-Physical Systems
The use of learning based components in cyber-physical systems (CPS) has created a gamut of possible avenues to use high dimensional real world signals generated from sensors like camera and LiDAR. The ability to process such signals can be largely attributed to the adoption of high-capacity function approximators like deep neural networks. However, this does not come without its potential perils. The pitfalls arise from possible over-fitting, and subsequent unsafe behavior when exposed to unknown environments. One challenge is that, in high dimensional input spaces it is almost impossible to experience enough training data in the design phase. What is required here, is an efficient way to flag out-of-distribution (OOD) samples that is precise enough to not raise too many false alarms. In addition, the system needs to be able to detect these in a computationally efficient manner at runtime. In this paper, our proposal is to build good representations for in-distribution data. We introduce the idea of a memory bank to store prototypical samples from the input space. We use these memories to compute probability density estimates using kernel density estimation techniques. We evaluate our technique on two challenging scenarios : a self-driving car setting implemented inside the simulator CARLA with image inputs, and an autonomous racing car navigation setting, with LiDAR inputs. In both settings, it was observed that a deviation from in-distribution setting can potentially lead to deviation from safe behavior. An added benefit of using training samples as memories to detect out-of-distribution inputs is that the system is interpretable to a human operator. Explanation of this nature is generally hard to obtain from pure deep learning based alternatives. Our code for reproducing the experiments is available at https:// yangy96/ interpretable_ood_detection.git  more » « less
Award ID(s):
Author(s) / Creator(s):
; ; ;
Date Published:
Journal Name:
ACMIEEE International Conference on CyberPhysical Systems
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Deep neural networks (DNNs) have achieved near-human level accuracy on many datasets across different domains. But they are known to produce incorrect predictions with high confidence on inputs far from the training distribution. This challenge of lack of calibration of DNNs has limited the adoption of deep learning models in high-assurance systems such as autonomous driving, air traffic management, cybersecurity, and medical diagnosis. The problem of detecting when an input is outside the training distribution of a machine learning model, and hence, its prediction on this input cannot be trusted, has received significant attention recently. Several techniques based on statistical, geometric, topological, or relational signatures have been developed to detect the out-of-distribution (OOD) or novel inputs. In this paper, we present a runtime monitor based on predictive processing and dual process theory. We posit that the bottom-up deep neural networks can be monitored using top-down context models comprising two layers. The first layer is a feature density model that learns the joint distribution of the original DNN’s inputs, outputs, and the model’s explanation for its decisions. The second layer is a graph Markov neural network that captures an even broader context. We demonstrate the efficacy of our monitoring architecture in recognizing out-of-distribution and out-of-context inputs on the image classification and object detection tasks. 
    more » « less
  2. Site description. This data package consists of data obtained from sampling surface soil (the 0-7.6 cm depth profile) in black mangrove (Avicennia germinans) dominated forest and black needlerush (Juncus roemerianus) saltmarsh along the Gulf of Mexico coastline in peninsular west-central Florida, USA. This location has a subtropical climate with mean daily temperatures ranging from 15.4 °C in January to 27.8 °C in August, and annual precipitation of 1336 mm. Precipitation falls as rain primarily between June and September. Tides are semi-diurnal, with 0.57 m median amplitudes during the year preceding sampling (U.S. NOAA National Ocean Service, Clearwater Beach, Florida, station 8726724). Sea-level rise is 4.0 ± 0.6 mm per year (1973-2020 trend, mean ± 95 % confidence interval, NOAA NOS Clearwater Beach station). The A. germinans mangrove zone is either adjacent to water or fringed on the seaward side by a narrow band of red mangrove (Rhizophora mangle). A near-monoculture of J. roemerianus is often adjacent to and immediately landward of the A. germinans zone. The transition from the mangrove to the J. roemerianus zone is variable in our study area. An abrupt edge between closed-canopy mangrove and J. roemerianus monoculture may extend for up to several hundred meters in some locations, while other stretches of ecotone present a gradual transition where smaller, widely spaced trees are interspersed into the herbaceous marsh. Juncus roemerianus then extends landward to a high marsh patchwork of succulent halophytes (including Salicornia bigellovi, Sesuvium sp., and Batis maritima), scattered dwarf mangrove, and salt pans, followed in turn by upland vegetation that includes Pinus sp. and Serenoa repens. Field design and sample collection. We established three study sites spaced at approximately 5 km intervals along the western coastline of the central Florida peninsula. The sites consisted of the Salt Springs (28.3298°, -82.7274°), Energy Marine Center (28.2903°, -82.7278°), and Green Key (28.2530°, -82.7496°) sites on the Gulf of Mexico coastline in Pasco County, Florida, USA. At each site, we established three plot pairs, each consisting of one saltmarsh plot and one mangrove plot. Plots were 50 m^2 in size. Plots pairs within a site were separated by 230-1070 m, and the mangrove and saltmarsh plots composing a pair were 70-170 m apart. All plot pairs consisted of directly adjacent patches of mangrove forest and J. roemerianus saltmarsh, with the mangrove forests exhibiting a closed canopy and a tree architecture (height 4-6 m, crown width 1.5-3 m). Mangrove plots were located at approximately the midpoint between the seaward edge (water-mangrove interface) and landward edge (mangrove-marsh interface) of the mangrove zone. Saltmarsh plots were located 20-25 m away from any mangrove trees and into the J. roemerianus zone (i.e., landward from the mangrove-marsh interface). Plot pairs were coarsely similar in geomorphic setting, as all were located on the Gulf of Mexico coastline, rather than within major sheltering formations like Tampa Bay, and all plot pairs fit the tide-dominated domain of the Woodroffe classification (Woodroffe, 2002, "Coasts: Form, Process and Evolution", Cambridge University Press), given their conspicuous semi-diurnal tides. There was nevertheless some geomorphic variation, as some plot pairs were directly open to the Gulf of Mexico while others sat behind keys and spits or along small tidal creeks. Our use of a plot-pair approach is intended to control for this geomorphic variation. Plot center elevations (cm above mean sea level, NAVD 88) were estimated by overlaying the plot locations determined with a global positioning system (Garmin GPS 60, Olathe, KS, USA) on a LiDAR-derived bare-earth digital elevation model (Dewberry, Inc., 2019). The digital elevation model had a vertical accuracy of ± 10 cm (95 % CI) and a horizontal accuracy of ± 116 cm (95 % CI). Soil samples were collected via coring at low tide in June 2011. From each plot, we collected a composite soil sample consisting of three discrete 5.1 cm diameter soil cores taken at equidistant points to 7.6 cm depth. Cores were taken by tapping a sleeve into the soil until its top was flush with the soil surface, sliding a hand under the core, and lifting it up. Cores were then capped and transferred on ice to our laboratory at the University of South Florida (Tampa, Florida, USA), where they were combined in plastic zipper bags, and homogenized by hand into plot-level composite samples on the day they were collected. A damp soil subsample was immediately taken from each composite sample to initiate 1 y incubations for determination of active C and N (see below). The remainder of each composite sample was then placed in a drying oven (60 °C) for 1 week with frequent mixing of the soil to prevent aggregation and liberate water. Organic wetland soils are sometimes dried at 70 °C, however high drying temperatures can volatilize non-water liquids and oxidize and decompose organic matter, so 50 °C is also a common drying temperature for organic soils (Gardner 1986, "Methods of Soil Analysis: Part 1", Soil Science Society of America); we accordingly chose 60 °C as a compromise between sufficient water removal and avoidance of non-water mass loss. Bulk density was determined as soil dry mass per core volume (adding back the dry mass equivalent of the damp subsample removed prior to drying). Dried subsamples were obtained for determination of soil organic matter (SOM), mineral texture composition, and extractable and total carbon (C) and nitrogen (N) within the following week. Sample analyses. A dried subsample was apportioned from each composite sample to determine SOM as mass loss on ignition at 550 °C for 4 h. After organic matter was removed from soil via ignition, mineral particle size composition was determined using a combination of wet sieving and density separation in 49 mM (3 %) sodium hexametaphosphate ((NaPO_3)_6) following procedures in Kettler et al. (2001, Soil Science Society of America Journal 65, 849-852). The percentage of dry soil mass composed of silt and clay particles (hereafter, fines) was calculated as the mass lost from dispersed mineral soil after sieving (0.053 mm mesh sieve). Fines could have been slightly underestimated if any clay particles were burned off during the preceding ignition of soil. An additional subsample was taken from each composite sample to determine extractable N and organic C concentrations via 0.5 M potassium sulfate (K_2SO_4) extractions. We combined soil and extractant (ratio of 1 g dry soil:5 mL extractant) in plastic bottles, reciprocally shook the slurry for 1 h at 120 rpm, and then gravity filtered it through Fisher G6 (1.6 μm pore size) glass fiber filters, followed by colorimetric detection of nitrite (NO_2^-) + nitrate (NO_3^-) and ammonium (NH_4^+) in the filtrate (Hood Nowotny et al., 2010,Soil Science Society of America Journal 74, 1018-1027) using a microplate spectrophotometer (Biotek Epoch, Winooski, VT, USA). Filtrate was also analyzed for dissolved organic C (referred to hereafter as extractable organic C) and total dissolved N via combustion and oxidation followed by detection of the evolved CO_2 and N oxide gases on a Formacs HT TOC/TN analyzer (Skalar, Breda, The Netherlands). Extractable organic N was then computed as total dissolved N in filtrate minus extractable mineral N (itself the sum of extractable NH_4-N and NO_2-N + NO_3-N). We determined soil total C and N from dried, milled subsamples subjected to elemental analysis (ECS 4010, Costech, Inc., Valencia, CA, USA) at the University of South Florida Stable Isotope Laboratory. Median concentration of inorganic C in unvegetated surface soil at our sites is 0.5 % of soil mass (Anderson, 2019, Univ. of South Florida M.S. thesis via methods in Wang et al., 2011, Environmental Monitoring and Assessment 174, 241-257). Inorganic C concentrations are likely even lower in our samples from under vegetation, where organic matter would dilute the contribution of inorganic C to soil mass. Nevertheless, the presence of a small inorganic C pool in our soils may be counted in the total C values we report. Extractable organic C is necessarily of organic C origin given the method (sparging with HCl) used in detection. Active C and N represent the fractions of organic C and N that are mineralizable by soil microorganisms under aerobic conditions in long-term soil incubations. To quantify active C and N, 60 g of field-moist soil were apportioned from each composite sample, placed in a filtration apparatus, and incubated in the dark at 25 °C and field capacity moisture for 365 d (as in Lewis et al., 2014, Ecosphere 5, art59). Moisture levels were maintained by frequently weighing incubated soil and wetting them up to target mass. Daily CO_2 flux was quantified on 29 occasions at 0.5-3 week intervals during the incubation period (with shorter intervals earlier in the incubation), and these per day flux rates were integrated over the 365 d period to compute an estimate of active C. Observations of per day flux were made by sealing samples overnight in airtight chambers fitted with septa and quantifying headspace CO_2 accumulation by injecting headspace samples (obtained through the septa via needle and syringe) into an infrared gas analyzer (PP Systems EGM 4, Amesbury, MA, USA). To estimate active N, each incubated sample was leached with a C and N free, 35 psu solution containing micronutrients (Nadelhoffer, 1990, Soil Science Society of America Journal 54, 411-415) on 19 occasions at increasing 1-6 week intervals during the 365 d incubation, and then extracted in 0.5 M K_2SO_4 at the end of the incubation in order to remove any residual mineral N. Active N was then quantified as the total mass of mineral N leached and extracted. Mineral N in leached and extracted solutions was detected as NH_4-N and NO_2-N + NO_3-N via colorimetry as above. This incubation technique precludes new C and N inputs and persistently leaches mineral N, forcing microorganisms to meet demand by mineralizing existing pools, and thereby directly assays the potential activity of soil organic C and N pools present at the time of soil sampling. Because this analysis commences with disrupting soil physical structure, it is biased toward higher estimates of active fractions. Calculations. Non-mobile C and N fractions were computed as total C and N concentrations minus the extractable and active fractions of each element. This data package reports surface-soil constituents (moisture, fines, SOM, and C and N pools and fractions) in both gravimetric units (mass constituent / mass soil) and areal units (mass constituent / soil surface area integrated through 7.6 cm soil depth, the depth of sampling). Areal concentrations were computed as X × D × 7.6, where X is the gravimetric concentration of a soil constituent, D is soil bulk density (g dry soil / cm^3), and 7.6 is the sampling depth in cm. 
    more » « less
  3. Deep-learning driven safety-critical autonomous systems, such as self-driving cars, must be able to detect situations where its trained model is not able to make a trustworthy prediction. This ability to determine the novelty of a new input with respect to a trained model is critical for such systems because novel inputs due to changes in the environment, adversarial attacks, or even unintentional noise can potentially lead to erroneous, perhaps life-threatening decisions. This paper proposes a learning framework that leverages information learned by the prediction model in a task-aware manner to detect novel scenarios. We use network saliency to provide the learning architecture with knowledge of the input areas that are most relevant to the decision-making and learn an association between the saliency map and the predicted output to determine the novelty of the input. We demonstrate the efficacy of this method through experiments on real-world driving datasets as well as through driving scenarios in our in-house indoor driving environment where the novel image can be sampled from another similar driving dataset with similar features or from adversarial attacked images from the training dataset. We find that our method is able to systematically detect novel inputs and quantify the deviation from the target prediction through this task-aware approach. 
    more » « less
  4. Contemporary autonomous vehicle (AV) benchmarks have advanced techniques for training 3D detectors, particularly on large-scale lidar data. Surprisingly, although semantic class labels naturally follow a long-tailed distribution, contemporary benchmarks focus on only a few common classes (e.g., pedestrian and car) and neglect many rare classes in-the-tail (e.g., debris and stroller). However, AVs must still detect rare classes to ensure safe operation. Moreover, semantic classes are often organized within a hierarchy, e.g., tail classes such as child and construction-worker are arguably subclasses of pedestrian. However, such hierarchical relationships are often ignored, which may lead to misleading estimates of performance and missed opportunities for algorithmic innovation. We address these challenges by formally studying the problem of Long-Tailed 3D Detection (LT3D), which evaluates on all classes, including those in-the-tail. We evaluate and innovate upon popular 3D detection codebases, such as CenterPoint and PointPillars, adapting them for LT3D. We develop hierarchical losses that promote feature sharing across common-vs-rare classes, as well as improved detection metrics that award partial credit to "reasonable" mistakes respecting the hierarchy (e.g., mistaking a child for an adult). Finally, we point out that fine-grained tail class accuracy is particularly improved via multimodal fusion of RGB images with LiDAR; simply put, small fine-grained classes are challenging to identify from sparse (lidar) geometry alone, suggesting that multimodal cues are crucial to long-tailed 3D detection. Our modifications improve accuracy by 5% AP on average for all classes, and dramatically improve AP for rare classes (e.g., stroller AP improves from 3.6 to 31.6)! Our code is available at this https URL. 
    more » « less
  5. Several recent research efforts have proposed Machine Learning (ML)-based solutions that can detect complex patterns in network traffic for a wide range of network security problems. However, without understanding how these black-box models are making their decisions, network operators are reluctant to trust and deploy them in their production settings. One key reason for this reluctance is that these models are prone to the problem of underspecification, defined here as the failure to specify a model in adequate detail. Not unique to the network security domain, this problem manifests itself in ML models that exhibit unexpectedly poor behavior when deployed in real-world settings and has prompted growing interest in developing interpretable ML solutions (e.g., decision trees) for “explaining” to humans how a given black-box model makes its decisions. However, synthesizing such explainable models that capture a given black-box model’s decisions with high fidelity while also being practical (i.e., small enough in size for humans to comprehend) is challenging. In this paper, we focus on synthesizing high-fidelity and low-complexity decision trees to help network operators determine if their ML models suffer from the problem of underspecification. To this end, we present TRUSTEE, a framework that takes an existing ML model and training dataset generate a high-fidelity, easy-to-interpret decision tree, and associated trust report. Using published ML models that are fully reproducible, we show how practitioners can use TRUSTEE to identify three common instances of model underspecification, i.e., evidence of shortcut learning, spurious correlations, and vulnerability to out-of-distribution samples. 
    more » « less