PaperlayerView exact revision on arXiv

Source-linked AI summary

MDRC: A Deployable State-Recovery Defense for Traffic Signal Control under Sensor Corruption

Mingyuan Li, Chunyu Liu, Xiao Liu, Yanna Jiang, Guangsheng Yu, Xu Wang, Wei Ni, Ren Ping Liu

arXiv:2609.27528v1cs.CR

TL;DR

Corrupted traffic-signal observations create a state-recovery problem, while robust TSC methods face cross-city transfer and latency challenges. MDRC inserts DDIM-based recovery before the controller and uses Reptile for cross-city initialization; it reports lower average travel time under attacks and sensor loss.

  • Problem

    Traffic signal control depends on sensing, but corrupted observations can degrade control, and robust RL methods face limited cross-city transfer, high overhead, and weak recovery under partial observability.

    Source: §I · p. 1

  • Method

    MDRC is a post-detection plug-in that reconstructs corrupted traffic states using DDIM and Reptile meta-learning for transferable recovery-model initialization.

    Source: §I · p. 2, §V · p. 4

  • Results

    MDRC lowers ATT by 6.77% on average under stochastic and policy-targeted attacks, and improves ATT by 12.75% on average over RobustLight under sensor loss.

    Source: C. Results · pp. 8–9

  • Takeaways & Limitations

    MDRC is presented as a controller-agnostic defense that reconstructs traffic states before control and improves performance across attack and sensor-failure settings.

    Source: §I · p. 2, §VIII · p. 13

Abstract

from arXiv · show

Traffic Signal Control (TSC) is a safety-critical cyber-physical system that relies on real-time sensing. Corrupted observations caused by adversarial perturbations or sensor failures can propagate from the sensing layer into the controller and degrade traffic efficiency. Existing robust Reinforcement Learning (RL)-based TSC methods often suffer from limited cross-city generalization, high inference latency, and weak recovery under partial observability. We present MDRC (Meta-Diffusion-based framework for Resilient traffic signal Control against adversarial attacks and sensor failures), a post-detection state-recovery defense inserted between sensing and control. MDRC reconstructs trustworthy traffic states before they are consumed by the controller. It combines Denoising Diffusion Implicit Models (DDIM) for efficient state recovery with Reptile meta-learning for a transferable initialization across cities. We provide an optimization-based view of the DDIM recovery dynamics and establish a recovery-error bound that separates score approximation, numerical discretization, and initialization mismatch. Across seven real-world-derived CityFlow benchmarks, MDRC reduces Average Travel Time by 6.77% under stochastic and policy-aware attacks and by 12.75% under structured sensor loss, while improving state-recovery fidelity. We further evaluate 3,600 seconds of real roadside measurements with 50% of detector channels disabled and integrate MDRC into a hardware-in-the-loop traffic-signal stack. Over a 9.16-hour run with 32,389 sensing/control cycles, the system achieves 99.79% decision availability, produces no out-of-plan recommendations, and requires approximately 38 ms of component-wise processing per one-second control interval.

I. INTRODUCTION

The paper frames resilient traffic signal control as recovering trustworthy states after sensing corruption, addressing transfer and latency gaps in existing robust learning. MDRC combines diffusion-based recovery with cross-city meta-learning, against a background of TSC, diffusion, and meta-learning methods.

I. INTRODUCTION

  • RobustLight’s per-city training, inference and adaptation overhead, and lack of explicit sensing-interface defense limit its fit for security-critical TSC deployment.MDRC instead targets post-detection recovery and forwards a reconstructed state to the original controller.Source: §I · p. 1
  • MDRC combines DDIM for fast state reconstruction with Reptile meta-learning for a transferable initialization that supports zero-shot deployment and few-shot adaptation across cities.The recovery layer is designed to work with RL-based and rule-based controllers without replacing them.Source: §I · p. 2

II. RELATED WORKS

  • Prior TSC work spans adaptive and RL control, while diffusion-based methods including DiffLight, RobustLight, and DMBP are described as slow and suboptimal at reconstruction.MDRC differs from classical model-based estimators such as Kalman filtering by targeting nonlinear cross-city distributions and structured loss without city-specific dynamics calibration.Source: §II · p. 2

III. PRELIMINARY

  • TSC states describe intersection movements and signal phases; DDIM speeds diffusion sampling with a strided schedule, while Reptile seeks an initialization near task-specific optima for rapid adaptation.Reptile is presented as a first-order, computationally efficient alternative in spirit to MAML.Source: §III · pp. 2–3

IV. THREAT MODEL AND ATTACK TAXONOMY

The threat model confines MDRC’s defense to post-detection corruption of observations before they reach the controller. It covers stochastic and policy-aware perturbations as well as sensor loss, while excluding attacks on downstream components.

Threat scope

  • The attacker targets the sensing-layer observation before the controller, while attacks on controller logic, parameters, rewards, training data, and actuation hardware are excluded.Recovery is assumed to be triggered by an upstream anomaly or sensor-health module, without an oracle clean state, attack label, or perturbation details.Source: §IV · pp. 3–4

Stochastic attacks

  • Stochastic attacks inject Gaussian noise or bounded uniform random noise into the observed state to mislead the controller.The experiments apply MDRC directly to attacked observations to isolate recovery robustness from detector accuracy.Source: §IV · p. 4

Adaptive attacks and sensor loss

  • Policy-aware black-box attacks use query access to craft perturbations that maximize action divergence or minimize Q-value, while sensor-availability attacks mask readings and create partial observability.The policy-aware attacker has no access to model parameters or gradients.Source: §IV · pp. 3–4

V. METHODS

MDRC inserts a post-detection recovery layer between sensing and control, reconstructing a state for the unchanged controller. DDIM performs iterative state-space recovery, while Reptile meta-learns a transferable model initialization across cities.

V. METHODS

  • After suspicious sensing is flagged, MDRC reconstructs a trustworthy state from the compromised observation and sends it to the original controller without changing control logic.The layer requires neither an oracle clean state nor an attack label, and is compatible with RL-based and rule-based TSC algorithms.Source: §V · p. 4

A. DDIM Recovery Dynamics and Reptile Meta-Learning

  • DDIM’s update is interpreted as noise-conditioned, time-varying descent in state space, with schedule-derived terms providing momentum-like effects and controlling sampling stochasticity.This optimization-like view explains the design without treating DDIM dynamics as Reptile-style parameter updates.Source: A. DDIM Recovery Dynamics and Reptile… · pp. 4–5
  • Reptile meta-learns recovery-model parameters across source-city tasks, while DDIM uses the resulting predictor to iteratively recover traffic states at inference time.The paper presents their connection as an interpretation, not an equivalence between Reptile and DDIM.Source: A. DDIM Recovery Dynamics and Reptile… · pp. 4–5, Figure 3 · p. 5

B. Framework Overview

MDRC combines DDIM-based state recovery with Reptile meta-learning to transfer a recovery model across cities. It uses denoising and repainting to reconstruct traffic states before they reach an unchanged controller.

B. Framework Overview

  • MDRC unifies DDIM state recovery with Reptile updates for cross-city traffic-signal control.The framework meta-learns a shared initialization across city tasks and uses diffusion recovery at inference time.Source: B. Framework Overview · p. 5
  • The diffusion learner trains offline on logged trajectories from multiple cities, with Reptile aggregating task-adapted parameters into a shared initialization.Each city is treated as a task; inner-loop updates use the diffusion loss and the outer update moves global parameters toward task-adapted parameters.Source: B. Framework Overview · p. 5
  • The recovery loss balances immediate and future timesteps by penalizing predicted-versus-true noise mismatch across a window of length N + M.The supplied passage describes this as balancing near-term and future prediction.Source: B. Framework Overview · p. 5
  • The trained model supports zero-shot reconstruction and can be fine-tuned on a small set of target-city sequences for few-shot adaptation.The model is frozen after source-city training; target-city fine-tuning uses the diffusion loss.Source: B. Framework Overview · p. 6
  • The recovery-error bound separates score-estimation error, numerical discretization error, and initialization mismatch.Reptile is not a separate bound term; meta-training can reduce score error indirectly by improving the transferred predictor.Source: B. Framework Overview · p. 6
  • DDIM denoising iteratively reconstructs the state, while repainting uses a known-versus-unknown mask to restore missing or damaged readings.The combined procedure reconstructs the full state and leaves the downstream controller unchanged.Source: B. Framework Overview · p. 6

C. Detailed Algorithms

The detailed algorithms implement cross-city diffusion meta-learning followed by optional target-city adaptation, alongside an iterative repaint procedure for masked observations.

C. Detailed Algorithms

  • Cross-City Diffusion Meta-Learning initializes shared parameters, adapts them on source-city tasks, aggregates parameter differences, and fine-tunes on the target task.The source-task inner update uses step size η, and the global update uses meta-learning rate µ.Source: C. Detailed Algorithms · pp. 6–7
  • The Repaint algorithm takes an initial signal estimate, context, and mask, then iterates K rounds with U inner steps to update known and unknown components.It samples Gaussian noise except at the first iteration and applies a stochastic update when needed.Source: C. Detailed Algorithms · p. 7
  • The repaint iterations produce a refined signal estimate for recovery under noisy or missing observations.The described procedure is aimed at inpainting or signal recovery.Source: C. Detailed Algorithms · p. 7

VI. REAL-WORLD SENSOR-FAILURE AND HARDWARE-IN-THE-LOOP VALIDATION

On real roadside measurements with half the detector channels unavailable, MDRC recovers more temporal structure than the missing-input trace, with few-shot adaptation further reducing reconstruction error. The hardware-in-the-loop evaluation also reports approximately 38 ms of component-wise processing per one-second control interval.

VI. REAL-WORLD SENSOR-FAILURE AND HARDWARE-IN-THE-LOOP VALIDATION

  • The roadside failure experiment uses genuine measurements with controlled channel dropout, preserving real temporal dynamics and cross-channel correlations.The setup provides controlled ground truth for measuring recovery quality.Source: §VI · p. 7
  • With 50% of detector channels unavailable, both zero-shot and few-shot MDRC recover more measured temporal structure than the unrecovered input, and few-shot adaptation further reduces reconstruction error.The trace contains 3,600 consecutive one-second observations from twelve movement-level channels; no future observations are used.Source: Figure 5 · p. 7, §VI · p. 7
  • Approximately 38 ms of component-wise processing occupies 3.8% of the one-second control interval.The reported components are approximately 28 ms for MDRC inference and 10 ms for the existing control path.Source: §VI · pp. 7–8
  • The hardware-in-the-loop setup verifies recovery within a physical sensing-to-control stack using roadside sensors and a commercial signal-controller interface.MDRC reconstructs unreliable channels before the unchanged downstream controller.Source: §VI · pp. 7–8

VII. EXPERIMENTS

The experiments evaluate MDRC across seven CityFlow datasets, traditional and RL-based controllers, and stochastic, adaptive, and sensor-failure settings. Results assess ATT, cross-city transfer, inference and training cost, and state-recovery fidelity.

C. Results

  • 6.77% lower Average Travel Time on average under stochastic and adaptive attacks, with gains reaching 26.18% under adaptive MinQ on JiNan1 with Advanced-CoLight.MDRC retains advantages across both attack types and tested strengths, though improvements are not universal across controller–attack pairs.Source: C. Results · p. 8, Table II · p. 9
  • 12.75% average improvement over RobustLight under sensor masking, rising to 52.13% in HangZhou1 with 50% sensor damage.At 50% masking, sensorW and sensorE fail jointly; MDRC is especially valuable in this harder partial-observability regime.Source: C. Results · p. 9, Table V · p. 10
  • Few-shot transfer with 100 target samples achieves the best overall performance across JiNan–HangZhou transfers, often recovering or surpassing clean-data performance.Larger-scale NewYork transfer improves by 6.92%, but MDRC underperforms RobustLight on Newyork1 under Mask.Source: C. Results · p. 10, Table VI · p. 11, Figure 7 · p. 12
  • 87.9% lower denoising runtime and 91.9% lower demasking runtime than RobustLight support online use; training fits on one RTX 4090 GPU and scales approximately linearly with cities and diffusion steps.The runtime comparisons are based on Advanced-CoLight.Source: Table IX · p. 12, C. Results · p. 12, Table X · p. 12
  • 10.57% average improvement under Kriging Missing and Random Missing accompanies generally stronger PSNR and MAE than RobustLight across noise and mask settings.The reported missingness conditions are 12.5% single-intersection-sensor failure and 12.5% full-intersection failure; improvements also persist with π-Light under Gaussian and U-rand perturbations.Source: C. Results · pp. 12–13
  • Meta-learning and DDIM serve distinct roles: MDRC achieves the lowest ATT across most tested settings, while DDIM makes inference substantially faster than DDPM-based alternatives.For few-shot JiNan-to-HangZhou transfer, a meta-learning rate of 0.1 and 100 diffusion steps give the best tested trade-off.Source: C. Results · pp. 11–12, Figure 8 · p. 13, Figure 9 · p. 13

VIII. CONCLUSION

MDRC is presented as a post-detection state-recovery defense for traffic signal control. The conclusion reports improved performance under several corruption settings and lower denoising and demasking times.

VIII. CONCLUSION

  • MDRC reconstructs trustworthy traffic states from corrupted observations and feeds them to the original controller.The paper frames it as a practical plug-in defense for scalable real-time urban traffic systems.Source: §VIII · p. 13
  • MDRC improves performance in unseen cities under stochastic attacks, adaptive policy-targeted attacks, and sensor failures.The conclusion also reports denoising and demasking times reduced by 87.9% and 91.9%, respectively.Source: §VIII · p. 13
  • Future work will explore joint detection-and-recovery pipelines and real-world deployment feedback under multi-intersection coordination.Source: §VIII · p. 13

ETHICS CONSIDERATIONS

The study addresses resilience of traffic-signal control, a cyber-physical infrastructure area with potential public-safety implications. Evaluations used controlled simulation or offline settings, measured roadside traces, and a hardware-in-the-loop setup that retained commercial-controller safety constraints.

ETHICS CONSIDERATIONS

  • The work examines corrupted or unavailable sensing in traffic-signal control, infrastructure whose malfunction could affect public safety.Source: ETHICS CONSIDERATIONS · p. 13
  • Adversarial and sensor-failure experiments used controlled simulation or offline settings rather than disrupting operational traffic infrastructure.The real-world evaluation used measured roadside traces with controlled channel masking; hardware-in-the-loop retained commercial-controller safety constraints.Source: ETHICS CONSIDERATIONS · p. 14
  • The study involved no human subjects, personally identifiable information, or sensitive user data.Source: ETHICS CONSIDERATIONS · p. 14

APPENDIX

The appendix describes reproducibility materials and provides a proof of a recovery-error bound for DDIM-Reptile. The proof discusses score approximation, numerical discretization, and initialization mismatch as distinct sources of error.

APPENDIX

  • The appendix provides source code, datasets, configuration files, evaluation scripts, and documentation for reproducing and validating results.The materials include preprocessing and evaluation scripts, environment configurations, and step-by-step setup, training, and evaluation instructions.Source: APPENDIX · p. 16
  • The authors state that ChatGPT 5.4 was used only for language editing, while conceptual content, technical descriptions, data analysis, and results were developed by the authors.Source: APPENDIX · p. 16

A. Proof of Theorem 1

  • Under the stated one-sided Lipschitz property, discretization error accumulates linearly with µ rather than being exponentially amplified.DDIM uses a finite sequence of N steps, with ∆t = T/N; the proof distinguishes this discretization interval from the Reptile outer-loop rate µ.Source: A. Proof of Theorem 1 · pp. 16–17
  • The recovery-error bound separates score approximation, discretization, and initialization terms.The proof defines the practical model distribution and combines continuous-time and discretization bounds; the displayed final expression includes Ldiff(θ), Cdisc · µ, and CinitW2(pT, qT).Source: A. Proof of Theorem 1 · pp. 16–17

B. Hyperparameters

The hyperparameter listing covers the UNet, diffusion training, and TSC RL agent configurations.

B. Hyperparameters

  • The UNet uses embedding dimension 64 and hidden dimension 256, with state/action dimensions set to 20/32 and 4, respectively.Source: B. Hyperparameters · p. 17
  • Diffusion training uses a non-Markovian step of 6, diffusion timestep 100, Adam learning rate 0.0003, and meta/single epoch settings of 25/90.Source: B. Hyperparameters · p. 17
  • The TSC RL agent uses discount factor 0.8, buffer capacity 12,000, batch size 20, and Adam learning rate 0.001.Its epsilon-greedy settings are initial 0.8, minimum 0.2, and decay 0.95.Source: B. Hyperparameters · p. 17

C. Sensitivity to Imperfect Fault Detection

The section tests MDRC when an upstream detector misses some faulty channels, using 50% sensor missingness and a no-recovery baseline. Across Jinan and Hangzhou, recovery improves traffic metrics at both tested recall levels, with larger benefits at 80% recall; additional experiments include New York-scale performance and image-quality and noise comparisons.

C. Sensitivity to Imperfect Fault Detection

  • The detector–MDRC evaluation uses 3,600-second closed-loop CityFlow rollouts, three paired trials, and mean ± standard deviation; Hangzhou includes zero-shot transfer and few-shot adaptation.ATT and mean waiting vehicles are minimized, throughput is maximized, and percentage changes use each network's corresponding 50%-missing no-recovery baseline.Source: C. Sensitivity to Imperfect Fault… · p. 17
  • 45.2% lower ATT, 37.8% higher throughput, and 68.6% fewer mean waiting vehicles in Jinan at 80% detection recall versus the unrecovered baseline.The comparison concerns 50% sensor missingness with imperfect missing-channel detection.Source: Table XII · p. 18, C. Sensitivity to Imperfect Fault… · p. 18
  • 18.0% lower ATT, 22.5% higher throughput, and 47.1% fewer waiting vehicles in zero-shot Hangzhou at 80% recall; few-shot adaptation yields 19.7%, 23.1%, and 50.4% improvements, respectively.Few-shot values are reductions in ATT and waiting and an increase in throughput.Source: Table XII · p. 18, C. Sensitivity to Imperfect Fault… · p. 18
  • 31.2% lower ATT, 25.4% higher throughput, and 51.8% fewer mean waiting vehicles in Jinan at 60% detection recall versus the unrecovered baseline.Source: Table XII · p. 18, C. Sensitivity to Imperfect Fault… · p. 18
  • 80% recall consistently outperforms 60%, but MDRC remains beneficial when 40% of missing channels go undetected; recovered states are not shown equivalent to clean observations.Detector precision is fixed at 100%, undetected missing channels remain zero-filled, and the reference is the same 50%-missing observation without recovery.Source: C. Sensitivity to Imperfect Fault… · pp. 17–18

D. Additional Experiments

Related papers

Loading 2609.27528v1…