<rss version="2.0" xmlns:atom="https://www.w3.org/2005/Atom">
  <channel>
    <title>Research in Progress (RIP)</title>
    <link>https://rip.trb.org/</link>
    <atom:link href="https://rip.trb.org/Record/RSS?s=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" rel="self" type="application/rss+xml" />
    <description></description>
    <language>en-us</language>
    <copyright>Copyright © 2026. National Academy of Sciences. All rights reserved.</copyright>
    <docs>http://blogs.law.harvard.edu/tech/rss</docs>
    <managingEditor>tris-trb@nas.edu (Bill McLeod)</managingEditor>
    <webMaster>tris-trb@nas.edu (Bill McLeod)</webMaster>
    <image>
      <title>Research in Progress (RIP)</title>
      <url>https://rip.trb.org/Images/PageHeader-wTitle-RIP.jpg</url>
      <link>https://rip.trb.org/</link>
    </image>
    <item>
      <title>Revolutionizing Coastal Infrastructure Durability with Pervious Concrete: A Cost-Effective, High-Performance Seawall</title>
      <link>https://rip.trb.org/View/2696019</link>
      <description><![CDATA[This project develops and validates a pervious concrete seawall system to reduce wave loads and mitigate scour-related degradation at lower cost and maintenance demand. The work integrates (i) high-fidelity finite element analysis for preliminary design, (ii) fabrication of pervious concrete with tuned porosity (15–35%) using durability-enhancing binders and engineered biochar, (iii) controlled wave flume experiments with instrumented specimens and backfill monitoring, and (iv) seawall design optimization accelerated by surrogate model and genetic algorithm.
To achieve the above mentioned integration, the research will proceed through a series of coordinated actions. First, the research team will build a high-fidelity finite element model, analyze the wave load in seawall, and achieve a preliminary design. Next, pervious concrete specimens with controlled porosity will be fabricated using the preliminary design and tested in a wave flume, which simulates real coastal conditions by generating programmable waves and measuring forces, displacements, and backfill scour behind the seawall. Finally, the team will apply a HyperNetwork, a neural architecture that dynamically generates predictive models, to estimate performance metrics such as energy dissipation and structural stability across different design configurations. The research team has rich experience in developing surrogate models for engineering applications and will complete building this HyperNetwork-based surrogate model in six months. This HyperNetwork will be used together with a genetic algorithm to search for Pareto-optimal designs that balance durability, hydraulic efficiency, and cost. This integrated approach ties together physical testing and advanced modeling to deliver practical, field-ready guidance with the objective of reducing wave-driven degradation and improving structural resilience in simple, cost-effective terms.
]]></description>
      <pubDate>Thu, 23 Apr 2026 16:44:25 GMT</pubDate>
      <guid>https://rip.trb.org/View/2696019</guid>
    </item>
    <item>
      <title>LLM-Orchestrated Multi-Layer Digital Twin Network for Cyber-Resilient Traffic Management</title>
      <link>https://rip.trb.org/View/2663602</link>
      <description><![CDATA[Modern connected traffic systems are increasingly vulnerable to cyberattacks capable of propagating rapidly across networked infrastructure, inducing unsafe signal states, traffic congestion, and emergency response delays. Existing anomaly detection approaches including statistical thresholds, rule-based Automated Traffic Signal Performance Measures (ATSPM) and Signal Phase and Timing (SPaT) flags, and classical machine-learning methods such as Isolation Forest and one-class Support Vector Machines operate on limited data modalities and cannot capture cross-layer cyber-physical interactions or operator intent, leaving critical detection gaps in complex attack scenarios.
This project develops a distributed multi-layer digital twin (DT) network for urban traffic systems, enhanced by a large language model (LLM) for context-aware cyber anomaly detection. The framework mirrors physical traffic behavior, cyber infrastructure status, and operational decision processes across a corridor of 4–6 interconnected intersections, enabling early identification of unsafe and malicious events that threaten roadway safety. Each traffic unit is represented by coordinated Physical, Cyber, and Decision Layers: the Physical Layer models real-time mobility and safety conditions using ATSPM, SPaT/MAP data, and detector activity; the Cyber Layer mirrors controller firmware, communication telemetry, and roadside unit status; and the Decision Layer captures operator actions, timing plan updates, and agency-defined safety constraints. A customized transportation-aware LLM ingests both structured telemetry and unstructured logs to generate semantic feature embeddings that capture cross-layer and cross-node dependencies.
A hybrid neural anomaly detection engine integrates Temporal Convolutional Networks (TCNs) to learn evolving traffic and communication behaviors over time with Graph Neural Networks (GNNs) to capture spatial interactions and coordinated disruptions across interconnected intersections. This TCN–GNN architecture enables accurate recognition of both localized cyber intrusions and distributed corridor-level attacks. Detection performance is validated against controlled cyber-attack scenarios—including SPaT spoofing, firmware manipulation, and malicious timing-plan overrides—executed within the DT environment. Upon anomaly detection, the LLM generates actionable mitigation suggestions, such as isolating compromised controllers or reverting to safe fallback signal plans, which are evaluated within the digital twin to ensure that every recommendation supports operational safety, low latency, and service continuity.
The 12-month effort proceeds in two phases: development and calibration of the distributed multi-layer DTs with LLM integration for context modeling, followed by anomaly detection training, validation, and mitigation evaluation. Target performance metrics include detection accuracy of at least 90%, false-positive rates below 10%, decision-support latency improvements of at least 30%, and safety metric improvements of at least 20%. The project delivers a pilot-ready prototype, detailed deployment guidelines, and an open software repository to accelerate adoption by transportation agencies. 
]]></description>
      <pubDate>Tue, 03 Feb 2026 15:28:33 GMT</pubDate>
      <guid>https://rip.trb.org/View/2663602</guid>
    </item>
    <item>
      <title>Use of Advanced Data Capture Tools on Measurements of Crack Lengths and Potholes for Estimates and Final Quantities</title>
      <link>https://rip.trb.org/View/2652613</link>
      <description><![CDATA[According to the Pavement Management Information System, the Kansas Department of Transportation (KSDOT) maintains 11,357 miles of pavement (counting miles in both directions of divided highways).  About 90% of this mileage is asphalt pavement. KSDOT’s contract maintenance work related to crack sealing, pot-hole patching, etc., is common for these pavements. 

The current measurement techniques use a measuring wheel, distance measuring instrument (DMI), etc. These techniques are highly susceptible to human errors and utilize considerable time and manpower. They also obstruct the traffic flow while conducting roadway measurements and putting the personnel at risk. The KSDOT idea submitted cites data collection via high-accuracy drone surveys but drone operations are restricted on KSDOT right of ways to prevent traveler distraction.   

Recent developments in camera technology and high-speed, high-resolution image capture at an affordable cost offer the opportunity for automation of measurements of crack lengths and potholes/patches for estimates and final quantities.  Example camera models include Vantrue S1 Pro 2.7K Front and Rear 5G WiFi Dash Cam, VIOFO Dash Cam Front and Rear 2K 1440P 60fps, Dash Cam Front and Rear - POFOTO 2.5K 1440P 60fps and 1080P 30fps Dash Camera, VIOFO A129 Plus Dash Cam 2K 1440P 60FPS GPS Wi-Fi Car Dash Camera with HDR and equivalent. These cameras all cost less than $250. 

The challenge lies in processing the images. However, with recent developments in artificial intelligence and machine learning, this problem can be resolved relatively quickly. One such algorithm for spatial pattern analysis is Convolutional Neural Networks (CNN), which have developed rapidly and have been applied in computer vision, natural language processing, and other fields. The convolutional neural network mimics the biological visual perception mechanism and can carry out supervised and unsupervised learning. However, traditional CNN has some drawbacks, like as the number of layers increases, the quality of the model decreases, ultimately leading to a decline in supervised learning accuracy. Thus, newer algorithms based on CNN have been developed that will be deployed in this study.]]></description>
      <pubDate>Tue, 13 Jan 2026 16:08:53 GMT</pubDate>
      <guid>https://rip.trb.org/View/2652613</guid>
    </item>
    <item>
      <title>Spatio Temporal Graph Learning for Real Time Pedestrian Exposure Estimation</title>
      <link>https://rip.trb.org/View/2640189</link>
      <description><![CDATA[Pedestrian crashes occur infrequently and are often underreported, which makes it difficult for agencies to rely only on crash records when assessing safety. Traditional Safety Performance Functions do not capture short term patterns or local context, and therefore cannot fully represent changes in pedestrian activity. This project will create a new framework that uses spatio temporal graph neural networks combined with statistical modeling to estimate pedestrian exposure across different locations and time periods. The research will draw from computer vision systems, Streetlight data, manual counts, roadway characteristics, land use, and travel related factors to produce high resolution exposure estimates.

The modeling framework will include two tiers. The first tier will use generalized linear mixed models to build a baseline exposure structure, while the second tier will apply deep learning methods to capture spatial spillover effects and temporal variation such as peak periods and seasonal changes. The results will help agencies identify areas with elevated pedestrian activity and evaluate how different roadway or land use conditions influence exposure. These data will support improved pedestrian safety analysis and guide the development of timely, evidence based interventions.]]></description>
      <pubDate>Thu, 11 Dec 2025 13:45:37 GMT</pubDate>
      <guid>https://rip.trb.org/View/2640189</guid>
    </item>
    <item>
      <title>Regional Disparities in Work Zone Crashes: Understanding Factors and Predictive Modeling for Targeted Safety Measures
</title>
      <link>https://rip.trb.org/View/2627354</link>
      <description><![CDATA[Roadway work zones play a vital role in maintaining and improving infrastructure, yet they often expose workers and drivers to dangerous situations, leading to concerning frequencies of occupational and traffic accidents in the United States. With over 700 fatalities and thousands of injuries annually attributed to work zone crashes, efforts to enhance safety have been hindered by the complexity and variability of contributing factors. The escalating fatalities, coupled with growing infrastructure demands in U.S. Department of Transportation Region 7—encompassing Missouri, Iowa, Nebraska, and Kansas—underscore the imperative to address underlying causes and improve work zone safety. This study aims to address this persistent issue by analyzing work zone crash data in Region 7 and comparing it with other regions to identify influential factors. By leveraging recurrent neural networks (RNNs) to develop region-specific predictive models, the research seeks to forecast crash occurrences and provide targeted insights for policymakers and transportation authorities. Ultimately, the research aims to deepen understanding of regional disparities in work zone crash dynamics, enabling effective resource prioritization and implementation of tailored safety measures. The development of predictive models using RNNs holds promise for enhancing proactive safety planning and resource allocation, ultimately contributing to a nationwide reduction in work zone crashes and advancing the overarching goal of improving road safety for workers and motorists.
]]></description>
      <pubDate>Wed, 19 Nov 2025 15:35:55 GMT</pubDate>
      <guid>https://rip.trb.org/View/2627354</guid>
    </item>
    <item>
      <title>Scooter-Share Travel Demand Forecast: A Context-Aware LSTM Recurrent Neural Network Approach</title>
      <link>https://rip.trb.org/View/2459122</link>
      <description><![CDATA[Shared micromobility has been popular in many cities in the U.S. The rise of shared micromobility brings significant operational challenges such as fleet management and demand forecasting. This project develops a Context-Aware Long Short-Term Memory (CALSTM) recurrent neural network to enhance the prediction of daily travel demand for scooter-sharing in Austin, Texas. The CALSTM model boosts prediction accuracy by integrating the impact of nearby points-of-interest (POIs) and daily weather conditions on scooter usage. It processes historical scooter-sharing demand and weather information through separate LSTM modules to extract temporal information. The outputs from these modules are combined through element-wise multiplication to establish temporal dependencies. Additionally, POI information is analyzed using a Multi-Layer Perceptron (MLP) to capture spatial dependencies. These spatial and temporal dependencies are then integrated by another MLP module to produce the forecast outputs. Case study experiments in Austin, TX, demonstrated that the CALSTM model significantly outperformed benchmark models, achieving improvements of 28% in Mean Absolute Error (MAE) and 19% in Root Mean Squared Error (RMSE) over traditional LSTM models. These results offer valuable insights for transportation planning and the enhancement of shared micromobility in urban settings.]]></description>
      <pubDate>Sat, 23 Nov 2024 11:10:42 GMT</pubDate>
      <guid>https://rip.trb.org/View/2459122</guid>
    </item>
    <item>
      <title>Physics-Informed Learning and Control of Connected and Autonomous Vehicles for Congestion Reduction</title>
      <link>https://rip.trb.org/View/2458989</link>
      <description><![CDATA[Building upon previous work in lane changing using physics-informed machine learning for autonomous vehicles, the goal of this project is to develop physics-informed machine learning and data-driven control-based tools for the combined longitudinal and lateral planning and control of connected and autonomous vehicles (CAVs). This research initiative holds the promise to have the following advantages: 1) Utilizing physics-informed machine learning as a tool can significantly enhance computational efficiency, which is beneficial for real-time control in complex scenarios; 2) Combining neural networks with physical models can greatly reduce over-reliance on data; 3) During the training phase of neural networks, any differentiable objective function and various constraints can be considered, allowing it to solve constrained multi-objective model predictive control problems without affecting computational speed. In addition, this project will design a lane-change decision-making module based on deep reinforcement learning and validate the congestion-reducing scheme using NGSIM data and SUMO simulations.]]></description>
      <pubDate>Thu, 21 Nov 2024 17:29:20 GMT</pubDate>
      <guid>https://rip.trb.org/View/2458989</guid>
    </item>
    <item>
      <title>Synthesis of Information Related to Airport Practices. Topic S07-06. Application of Large Language Models in Enhancing Airport Passenger Experience



</title>
      <link>https://rip.trb.org/View/2458785</link>
      <description><![CDATA[Airports today face an increasingly complex landscape characterized by a few challenges, such as operational inefficiencies, fluctuating passenger demands, and the ever-growing need to enhance the customer experience. These challenges are exacerbated by the dynamic nature of air travel, which demands rapid adaptation to changing conditions, whether due to unexpected surges in passenger numbers, shifts in travel patterns, or the need to respond to sudden disruptions. Traditional approaches to managing these issues relied on human inputs, which can cause delays, increased operational costs, and suboptimal passenger experiences. Large Language Models (LLMs) have the ability to interpret diverse forms of user textual inputs and generate contextual responses on demand. By processing vast amounts of data in real-time and efficiently delivering responses, LLMs have the potential to revolutionize the way airports interact with passengers and enhance operational efficiencies. Recognizing this potential, some airports have begun integrating this technology into their operations. However, it remains unclear how the broader airport industry can fully leverage LLMs. The potential costs and risks associated with adopting LLMs have yet to be thoroughly explored. 

The objective of this synthesis is to document how Large Language Models (LLMs) are being incorporated  into the airport environment to enhance passenger experience. The audience for this synthesis are airport practitioners that are responsible for passenger experience and those involved in technology evaluation.]]></description>
      <pubDate>Mon, 18 Nov 2024 20:38:23 GMT</pubDate>
      <guid>https://rip.trb.org/View/2458785</guid>
    </item>
    <item>
      <title>Low-cost Real-Time Learning-based Localization for Autonomous Vehicles</title>
      <link>https://rip.trb.org/View/2440024</link>
      <description><![CDATA[A major operational expense for an autonomous vehicle (AV) is with capturing, processing, and updating high-definition maps to localize itself when driving. To safely navigate, AVs need to know where they are on a given map to determine their trajectory to the next waypoint. Precise localization is a challenge in Global Positioning System (GPS)-denied areas such as dense urban corridors and motion tracking experiences dropouts in large open spaces such as rural highways. Classic localization algorithms are iterative, and their performance relies on direct feature matching between the stored map and the current sensor observations. This makes them prone to errors in large open spaces which have few distinguishing surface features. They are expensive to run in the vehicle as they account for a large share of the computation cost and power consumption. Better accuracy and faster localization directly improve AV safety as they navigate around people and cluttered environments.
 
This project will develop an AV localization service that is low-cost, accurate, and can operate in real-time in any AV at a fraction of the computation and power budget of current approaches. In 2023-24 the research team developed the preliminary version of this localization approach using a specific type of neural networks (i.e. invertible neural networks) to compress the map and lookup the vehicle’s pose efficiently. The team demonstrated the accuracy and cost to operate on 1/10th-scale vehicles and benchmarked the performance using localization datasets to benchmark the performance. In 2024-25, the team will undertake the real-world evaluation on real AVs with their deployment partner, The Autoware Foundation. The team will focus on localization of an electric autonomous goods and person cart for intralogistics for indoor and outdoor navigation. The outcome of this work will result in a portable and easy-to-use localization system for Safety21 projects.
 
Technical details: AV localization is the problem of finding a robot’s pose using a map and sensor measurements, like LiDAR scans and camera images. However, finding injective mappings between measurements and poses is difficult because sensor measurements from multiple distant poses can be similar. To solve this ambiguity, Monte Carlo Localization, the widely adopted method, uses random hypothesis sampling and sensor measurement updates to infer the pose. Other common approaches are to use Bayesian filtering or to find better distinguishable global descriptors on the map. Recent developments in localization research usually propose better measurement models or feature extractors within these frameworks. In this project, the team proposes a radically new approach to frame the localization problem as an ambiguous inverse problem and solve it with an invertible neural network (INN). The team has recently demonstrated that INNs are naturally suitable for the localization problem with many benefits, in terms of high accuracy (within 0.25m for city-scale maps), high-speed operation (>150Hz) and operate on low-cost embedded system hardware. The team will demonstrate this on point-cloud and camera datasets with evaluation on indoor and outdoor localization benchmarks, and also deploy it on real autonomous vehicles around the 23-acre Pennovation campus to show real-time and scalable operation. ]]></description>
      <pubDate>Sun, 13 Oct 2024 09:38:39 GMT</pubDate>
      <guid>https://rip.trb.org/View/2440024</guid>
    </item>
    <item>
      <title>A Pilot Experimental Project for Predicting Pedestrian Flows using Computer Vision and Deep Learning</title>
      <link>https://rip.trb.org/View/2440260</link>
      <description><![CDATA[Walking for transportation, health, and pleasure is an essential part of people’s lives in most cities. Knowing where people linger, the destinations that attract them, and how those places are accessed could assist in optimizing business locations and providing better security. In addition, predicting and sharing congestion times and locations (perhaps in real-time as in Waze for cars) could also provide useful information to travelers who can then choose appropriate travel routes and improve travel efficiency. Yet, far less is known about the spatial and temporal variations in pedestrian volumes than is known about vehicular movement. While pedestrian route choice has been an active area of research, few studies have attempted to predict pedestrian flows from unbiased pedestrian count data. Pedestrian route choice models assume that people choose their walking routes based on their perceived path attributes. Statistical path choice models identify people’s behavior related to route attributes on the selected path. These models hypothesize that the fundamental utility attribute is path length or travel time, which pedestrians generally minimize. These models also consider that people are willing to deviate to longer routes if the preferred path is comparatively safe, comfortable, and aesthetically pleasing. Yet, these models are inefficient for pedestrian traffic planning since they require prohibitive amounts of information about individual walkers. The research team develops a graph convolutional network model (GCN) based only on pedestrian counts at various intersections and segments to predict pedestrian traffic flows.
]]></description>
      <pubDate>Thu, 10 Oct 2024 16:01:37 GMT</pubDate>
      <guid>https://rip.trb.org/View/2440260</guid>
    </item>
    <item>
      <title>Estimation of logistic transportation system performance under extreme weather condition: A data-driven approach</title>
      <link>https://rip.trb.org/View/2422927</link>
      <description><![CDATA[This is the year two effort of the 2-year project. The objective is to utilize a data-driven
approach (e.g., Machine Learning and Deep Learning) to robustly, proactively, and trustworthily estimate the impact of impending extreme weather events (e.g., tropical cyclones) on key logistical infrastructure elements, including ports and highways. Beyond the first year’s research, the study team plans to broaden its objectives from focusing on the port system (multiple points) to encompassing the interconnected network of the highway system. The highway network will be conceptualized as a graph in which intersections and roads are nodes and edges, respectively. A Graph Neural Network (GNN) equipped with temporal attention will be utilized to estimate the impact of tropical cyclones on highway operational performance, leveraging its ability to process complex spatial and temporal dependencies within the highway network.]]></description>
      <pubDate>Thu, 29 Aug 2024 14:21:24 GMT</pubDate>
      <guid>https://rip.trb.org/View/2422927</guid>
    </item>
    <item>
      <title>Subsurface Seismic Imaging Using Full-Waveform Inversion and Physics-Informed Neural Networks</title>
      <link>https://rip.trb.org/View/2387190</link>
      <description><![CDATA[Roadway subsidence presents a significant challenge in the maintenance and safety of transportation infrastructure. This localized downward movement of the ground surface is largely due to buried low-velocity anomalies, such as highly compressible soft clay or loose sand zones, voids, and abandoned mine workings. Subsidence not only compromises the integrity of the road surface but also poses a considerable risk to the safety of the traveling. The ability to effectively assess and address this geohazard is, therefore, a crucial aspect of transportation system management. The early identification of subsurface anomalies is key to mitigating risks associated with roadway subsidence. By detecting potential hazards before they manifest as surface deformations, remedial actions can be undertaken to prevent extensive damage or catastrophic collapse of the roadway. This proactive approach to roadway maintenance ensures the continuous safety and efficiency of transportation routes, thereby minimizing disruptions and potential hazards to the public. The overall objective of this research is to integrate Physics-Informed Neural Networks with full-waveform inversion to solve the elastic wave equation in heterogeneous geomaterials and invert subsurface low-velocity anomalies.]]></description>
      <pubDate>Tue, 04 Jun 2024 14:11:56 GMT</pubDate>
      <guid>https://rip.trb.org/View/2387190</guid>
    </item>
    <item>
      <title>Real Time Classification of Vehicle Types and Modes using Image Analysis and Data Fusion</title>
      <link>https://rip.trb.org/View/2353426</link>
      <description><![CDATA[Description: The goal of this project is to conduct a feasibility study on the development of software and selection of hardware that will measure multiple transportation modes and classify vehicles by their Federal Highway Administration (FHWA) classification. The research team will install several combined computer/camera systems to monitor the multi-modal traffic in the proximity of the University of South Carolina campus. This area has multiple transportation users, including pedestrians, mopeds, bicycles, motorcycles, passenger cars, trucks, trains and buses. Along with the video data, additional traffic collection sources such as pneumatic tubes and Bluetooth will be used. Multiple cameras will allow three dimensional data of the environment to be constructed in the software. The video data will be combined with other data using statistical updating methods (Bayesian) to produce final multi-modal traffic information. We will also explore counting traffic in non-typical locations, such as counting the number of pedestrians in/outside of cross walks in the roadway or pedestrians crossing stopped trains.

Intellectual Merit: (1) Image subtractions from successive images will be used to identify objects in the area of interest. (2) A discriminate function based on the object geometry and image texture will be used to classify objects. (3) The development of the object discriminant function as well as utilization of digital image correlation or other video object motion determination approaches will be the major contribution of this research.

Broader Impacts: Broader Impacts: The collection and analysis of integrated multimodal movement of people and goods will provide transportation planners with better quantitative information about the existing system. Beyond providing raw counts, an integrated video based system could provide information about unsafe practices of pedestrians and moped users that could be used to improve safety for these users.]]></description>
      <pubDate>Mon, 25 Mar 2024 15:48:19 GMT</pubDate>
      <guid>https://rip.trb.org/View/2353426</guid>
    </item>
    <item>
      <title>Developing a Portable Railroad Crossing Monitoring System Based on Artificial Intelligence and Image Processing Technology</title>
      <link>https://rip.trb.org/View/2335144</link>
      <description><![CDATA[The objective of this proposal is to create a cost-effective, field-deployable system capable of identifying, counting, and categorizing a diverse range of objects, including vehicles, pedestrians, and other foreign obstructions, at railroad grade crossings. This system also aims to supply crucial data for collision warnings, as well as inform future traffic management and urban planning initiatives. The cornerstone of a successful intelligent railroad grade crossing monitoring system lies in precise object detection, counting, and classification capabilities. To achieve this, the research team proposes the development of a specialized deep neural network (DNN) augmented with a custom detection algorithm. This network will operate in conjunction with an edge computing platform and commercially available cameras to identify potential hazards at grade crossings in real-time. Powered by batteries for enhanced portability, the system can be strategically deployed at specific crossings based on situational needs. Beyond basic detection, the proposed system will also excel in object classification, segregating detected objects into distinct categories such as pedestrian, vehicle, tree, or package. This nuanced classification will enable a shift from current “passive” warning mechanisms to a more “proactive” traffic management strategy. By recognizing and categorizing potential hazards, local agencies will be better equipped to make informed decisions for urban development, thereby mitigating trespassing risks by targeting their sources directly.]]></description>
      <pubDate>Tue, 06 Feb 2024 17:21:20 GMT</pubDate>
      <guid>https://rip.trb.org/View/2335144</guid>
    </item>
    <item>
      <title>Privacy-preserving Cyber-Safe Machine Learning Models for Traffic Forecasting</title>
      <link>https://rip.trb.org/View/2301355</link>
      <description><![CDATA[Traffic congestion not only disrupts transportation systems but also poses significant 
cybersecurity risks, particularly in protecting sensitive driver data. Accurate traffic forecasting is essential to mitigate these disruptions, yet traditional approaches often compromise privacy and expose critical data to potential threats. Leveraging raw driver data without adequate security measures can lead to privacy breaches and cybersecurity vulnerabilities. We propose a novel approach utilizing advanced cryptographic techniques on encrypted data to predict traffic conditions without compromising individual driver location privacy. Our technique leverages cutting-edge quadratic functional encryption to construct a semi-private neural network. This network consists of two parts: a private section that handles encrypted location reports, aggregates them, and generates outputs for subsequent non-encrypted layers. These layers enhance forecasting performance while minimizing storage, computation, and bandwidth requirements. Functional encryption serves as the cryptographic foundation for achieving these goals. First, we establish a comprehensive, spatiotemporal route format with unique encrypted identifiers for each 
route segment. Second, we analyze the impact of privacy levels on accuracy, acknowledging the trade-off between privacy and precision. This project also integrates an adaptable framework to account for dynamic traffic changes due to accidents, events, or weather, with a strong focus on maintaining data security throughout. We recognize that privacy-preserving techniques may increase user-side computation and communication overhead, potentially impacting system performance. Therefore, we will investigate the trade-offs between accuracy, resource allocation, and cybersecurity. We request a one-year extension to thoroughly explore encryption techniques and their impact on performance and privacy, ensuring a robust solution for privacy-preserving traffic forecasting.
]]></description>
      <pubDate>Mon, 04 Dec 2023 17:03:07 GMT</pubDate>
      <guid>https://rip.trb.org/View/2301355</guid>
    </item>
  </channel>
</rss>