Life of Research
Appearance
Preprint Open access CC BY 4.0

A Recovered Average Is Not a Defended Corridor: What Adversarial Training Buys Spatiotemporal Traffic Forecasters, Why the Case That Static Methods Fail Rests on Node Selection While the Evidence of Failure Comes From Attack Shape, and the Evaluation That Would Tell Them Apart

  1. SONYTECH

Preprint

DOI10.5281/zenodo.24430423 LOR-00260

Download PDF (379.2 KB) Cite

Cite

A Recovered Average Is Not a Defended Corridor: What Adversarial Training Buys Spatiotemporal Traffic Forecasters, Why the Case That…

Read

Abstract

(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0.

Life of Research Spatiotemporal graph networks forecast speed and flow across road sensor networks, and a line of work since 2021 shows that a few perturbed sensors can wreck those forecasts. Whether adversarial training repairs the damage is disputed. One position, stated most sharply by the reinforced dynamic adversarial training (RDAT) paper at KDD 2023, holds that adversarial training designed for static domains does not carry over to traffic, because the attacker picks different victim sensors at each moment, and that defence needs a learned, time-varying choice of training nodes. A second position, from attack papers, time-series robustness studies and a 2025 study of forecasting foundation models, reports that ordinary adversarial training, or for the foundation models lightweight latent adversarial fine-tuning, already recovers most of the lost accuracy. A third, from a 2026 re-evaluation, holds that adversarial training tuned to norm-bounded perturbations barely transfers to targeted, congestion-shaped attacks, and that even training on the attack itself loses ground when the attack's reach changes. In this analysis the author reviews and synthesises published results to show that the three positions are answering different questions under one word. A "dynamic attack" can vary in at least five things: which sensors it touches over time, the shape and size of the perturbation, the objective, the attacker's knowledge, and the metric used to score it. The dynamic-domain thesis concerns only the first. In RDAT's own main tables, adversarial training with randomly chosen nodes, the baseline RDAT presents as traditional adversarial training and elsewhere labels a dynamic strategy, closes about 93 percent of the attack-induced error on both benchmarks with Graph WaveNet, including against the time-dependent attack. Against that attack the learned selector adds less than two points of the gap, within run-to-run spread; under the other four attacks it stays within the spread on PeMS-D4 but adds 3.3 to 3.8 points on PeMS-BAY, slightly more than AT's seed spread. In the appendix, under the time-dependent attack, the ordering of the two methods flips with architecture while both close only 52 to 71 percent. Node selection is also the smaller lever in the attack literature's own ablations. The failures that are large come from changes in perturbation shape, target and metric. The paper consolidates the measured values in a table and a figure, states a procedure for crediting a robustness gain to a defence's dynamic component, and names the evaluations that would settle what remains open.

Keywords

Files (1)

Cite this record

Mahendrakar, P. (2026). A Recovered Average Is Not a Defended Corridor: What Adversarial Training Buys Spatiotemporal Traffic Forecasters, Why the Case That Static Methods Fail Rests on Node Selection While the Evidence of Failure Comes From Attack Shape, and the Evaluation That Would Tell Them Apart [Preprint]. Life of Research. https://doi.org/10.5281/zenodo.24430423

APA 7th edition
Export to a reference manager
BibTeX
@misc{mahendrakar2026recovered,
  author       = {Mahendrakar, Pranay},
  title        = {{A Recovered Average Is Not a Defended Corridor: What Adversarial Training Buys Spatiotemporal Traffic Forecasters, Why the Case That Static Methods Fail Rests on Node Selection While the Evidence of Failure Comes From Attack Shape, and the Evaluation That Would Tell Them Apart}},
  year         = {2026},
  month        = nov,
  date         = {2026-11-22},
  publisher    = {Life of Research},
  howpublished = {Preprint},
  doi          = {10.5281/zenodo.24430423},
  url          = {https://lifeofresearch.com/preprints/260},
  abstract     = {(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Life of Research Spatiotemporal graph networks forecast speed and flow across road sensor networks, and a line of work since 2021 shows that a few perturbed sensors can wreck those forecasts. Whether adversarial training repairs the damage is disputed. One position, stated most sharply by the reinforced dynamic adversarial training (RDAT) paper at KDD 2023, holds that adversarial training designed for static domains does not carry over to traffic, because the attacker picks different victim sensors at each moment, and that defence needs a learned, time-varying choice of training nodes. A second position, from attack papers, time-series robustness studies and a 2025 study of forecasting foundation models, reports that ordinary adversarial training, or for the foundation models lightweight latent adversarial fine-tuning, already recovers most of the lost accuracy. A third, from a 2026 re-evaluation, holds that adversarial training tuned to norm-bounded perturbations barely transfers to targeted, congestion-shaped attacks, and that even training on the attack itself loses ground when the attack's reach changes. In this analysis the author reviews and synthesises published results to show that the three positions are answering different questions under one word. A "dynamic attack" can vary in at least five things: which sensors it touches over time, the shape and size of the perturbation, the objective, the attacker's knowledge, and the metric used to score it. The dynamic-domain thesis concerns only the first. In RDAT's own main tables, adversarial training with randomly chosen nodes, the baseline RDAT presents as traditional adversarial training and elsewhere labels a dynamic strategy, closes about 93 percent of the attack-induced error on both benchmarks with Graph WaveNet, including against the time-dependent attack. Against that attack the learned selector adds less than two points of the gap, within run-to-run spread; under the other four attacks it stays within the spread on PeMS-D4 but adds 3.3 to 3.8 points on PeMS-BAY, slightly more than AT's seed spread. In the appendix, under the time-dependent attack, the ordering of the two methods flips with architecture while both close only 52 to 71 percent. Node selection is also the smaller lever in the attack literature's own ablations. The failures that are large come from changes in perturbation shape, target and metric. The paper consolidates the measured values in a table and a figure, states a procedure for crediting a robustness gain to a defence's dynamic component, and names the evaluations that would settle what remains open.},
  keywords     = {traffic forecasting, spatiotemporal graph neural networks, adversarial training, adversarial robustness, threat models, time-series forecasting, sensor spoofing, evaluation validity},
  language     = {en},
  copyright    = {Creative Commons Attribution 4.0}
}

Using Zotero or Mendeley? Their browser connector saves this record from this page, with its PDF when there is one. On search and browse pages it lists every result so you can save many at once.

Guides How to cite a preprint Preprints vs journal articles What CC BY 4.0 allows