Research map / Adversarial robustness and security / Attacks on text and graphs

RobustLight: Improving Robustness via Diffusion Reinforcement Learning for Traffic Signal Control

ICML 2025 · Adversarial robustness and security › Attacks on text and graphs · approach: Attacks and defenses in sequential settings

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RobustLight, a plug-and-play online diffusion-based state recovery framework (DSI agent with denoise and repaint algorithms) that restores adversarially attacked or missing sensor data so existing traffic signal control methods keep performing well.

Problem

Real-time traffic signal control (TSC) degrades under sensor adversarial noise and missing/damaged sensor data, and existing defenses handle only one attack type or rely on offline data.

Approach

An online actor-critic style dynamic state infilling agent trains a conditional U-Net diffusion model from the TSC replay buffer; at test time, a denoise algorithm and a RePaint-style inpainting algorithm restore attacked or masked states before they reach the unchanged TSC controller.

Key results

• RobustLight improves Advanced-CoLight average travel time under Gaussian noise (scale 3.5) on HangZhou1 from 480.38 to 327.98 seconds.
• The paper claims up to 50.43% improvement in average travel time under attacks and missing data.
• Under 50% sensor damage on JiNan1, Advanced-CoLight ATT drops from 732.37 to 680.37 with RobustLight, a modest gain.

Measured effects reported in the paper

Surprising findings

• Under 50% sensor damage in some settings RobustLight gives little or no improvement, and for some methods (e.g., HangZhou1 Advanced-CoLight 50% damage) results are not strictly better than baseline.

Limitations

• Slow inference from many diffusion steps, especially in large-scale networks with many intersections.
• Assumes attacks are pre-detected; detection accuracy affects performance (tested at 80% and 60% detection rates).
• Evaluated only in CityFlow simulation with synthetic attacks.

Experimental setup

Datasets: JiNan1 JiNan2 JiNan3 HangZhou1 HangZhou2 Newyork1 Newyork2

Benchmarks: CityFlow JiNan CityFlow HangZhou CityFlow New York

Metrics: average travel time reconstruction error E_denoise PSNR MAE throughput

Baselines: FixedTime MaxPressure Advanced-MaxPressure CoLight Advanced-CoLight Advanced-MpLight DiffLight LSTM

Builds on: Denoising diffusion probabilistic models RePaint Classifier-free guidance CoLight Advanced-MPLight

Closest papers by meaning

• DiffLight: A Partial Rewards Conditioned Diffusion Model for Traffic Signal Control with Missing Data (NeurIPS 2024 · similarity 0.90)
• π-Light: Programmatic Interpretable Reinforcement Learning for Resource-Limited Traffic Signal Control (AAAI 2024 · similarity 0.87)
• VLMLight: Safety-Critical Traffic Signal Control via Vision-Language Meta-Control and Dual-Branch Reasoning Architecture (NeurIPS 2025 · similarity 0.87)
• Prompt to Transfer: Sim-to-Real Transfer for Traffic Signal Control with Prompt Learning (AAAI 2024 · similarity 0.86)
• Adversarial Diffusion for Robust Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
• RLFTSim: Realistic and Controllable Multi-Agent Traffic Simulation via Reinforcement Learning Fine-Tuning (CVPR 2026 · similarity 0.85)