Adversarial Attacks on Downstream Weather Forecasting Models: Application to Tropical Cyclone Trajectory Prediction
Organizations: Michigan State University, East Lansing, MI 48824 USA
Abstract
Deep learning-based weather forecasting (DLWF) models leverage past weather observations to generate future forecasts, supporting a wide range of downstream applications, including tropical cyclone (TC) prediction. In this paper, we investigate the vulnerability of these models to adversarial attacks, where subtle perturbations to upstream forecasts can alter downstream TC trajectory predictions. Crafting perturbed inputs to steer downstream outputs toward attacker-specified trajectories is a challenging problem for several reasons. First, conventional TC detection systems are opaque, non-differentiable black boxes, making standard gradient-based attacks infeasible. Second, the extreme rarity of TC events leads to a severe class imbalance problem, making it difficult to develop attack methods that produce realistic-looking cyclone paths aligned with target trajectories. To overcome these limitations, we propose Cyc-Attack, a novel method for perturbing the forecasts of DLWF models to generate adversarial trajectories. The proposed method uses a differentiable surrogate model to approximate the TC detector's output, enabling the application of gradient-based attacks. Cyc-Attack also employs a skewness-aware loss function with a kernel dilation strategy to address the severe class imbalance problem. Distance-based gradient weighting and regularization strategies are also introduced to constrain perturbations to eliminate unrealistic-looking TC paths. Our experiments show that Cyc-Attack achieves a higher true positive rate in matching the target trajectories, along with lower false alarm rates and stealthier perturbations than conventional attack methods. We further demonstrate that the adversarial upstream forecasts generated by Cyc-Attack can be leveraged as initial guidance to train an end-to-end model for learning adversarial input samples that would induce target trajectories.
Figures & tables
| TNR | FNR | FPR | TPR | |
|---|---|---|---|---|
| 0 | 0.9933 | 0.0104 | 0.0067 | 0.9896 |
| 1 | 0.9999 | 0.4135 | 0.0001 | 0.5865 |
| 2 | 0.9998 | 0.1869 | 0.0002 | 0.8131 |
| 3 | 0.9993 | 0.0692 | 0.0007 | 0.9308 |
| 5 | 0.9986 | 0.0606 | 0.0014 | 0.9394 |
| Method | TC1 Location-level | TC1 Trajectory-level | TC1 Closeness | |||
|---|---|---|---|---|---|---|
| FPR | FNR | TPR | DR | FAR | ||
| ALA | 0.0001 ( 0.0000) | 0.8218 ( 0.0293) | 0.1782 ( 0.0293) | 0.3197 ( 0.0864) | 0.1393 ( 0.0572) | 0.0459 ( 0.0002) |
| TAAOWPF | 0.0002 ( 0.0000) | 0.8640 ( 0.0263) | 0.1360 ( 0.0263) | 0.1038 ( 0.0414) | 0.5991 ( 0.0610) | 0.1753 ( 0.0010) |
| AOWF | 0.0001 ( 0.0000) | 0.7971 ( 0.0322) | 0.2029 ( 0.0322) | 0.2977 ( 0.0749) | 0.1527 ( 0.0585) | 0.0450 ( 0.0007) |
| Cyc-Attack | 0.0001 ( 0.0000) | 0.7001 ( 0.0352) | 0.2999 ( 0.0352) | 0.5570 ( 0.0793) | 0.0537 ( 0.0345) | 0.0004 ( 0.0000) |
| Cyc-Attack- d | 0.0001 ( 0.0000) | 0.8476 ( 0.0252) | 0.1524 ( 0.0252) | 0.2804 ( 0.0841) | 0.0935 ( 0.0573) | 0.0003 ( 0.0000) |
| Dataset TC1 | Dataset TC2 | |||||||
|---|---|---|---|---|---|---|---|---|
| Metric | Cyc-Attack | ALA | Cyc-Attack | ALA | ||||
| FPR | 0.0001 | 0.0001 | 0.376 | 0.7078 | 0.0004 | 0.0007 | 0.0241 | |
| FNR | 0.6597 | 0.8292 | 0.3963 | 0.8275 | 0.0012 | |||
| TPR | 0.3403 | 0.1708 | 5.915 | 0.6037 | 0.1725 | 4.677 | 0.0012 | |
| DR | 0.5875 | 0.3167 | 3.638 | 0.0005 | 0.7250 | 0.1375 | 4.222 | 0.0039 |
| FAR | 0.0333 | 0.0917 | 0.0909 | 0.0875 | 0.7250 | 0.0009 | ||
| Method | FPR | FNR | TPR | DR | FAR | |
|---|---|---|---|---|---|---|
| Cyc-At-e2e | 0.0002 | 0.6202 | 0.3798 | 0.6471 | 0.2353 | 0.0502 |
| ALA-e2e | 0.0003 | 0.9880 | 0.0120 | 0.0400 | 0.6400 | 0.4337 |
| Cyc-At-ds | 0.0004 | 0.4447 | 0.5553 | 0.7500 | 0.0833 | N/A |
| ALA-ds | 0.0007 | 0.7380 | 0.2620 | 0.2222 | 0.5926 | N/A |
| Variable | TC1 Ground-truth Range | TC2 Adversarial Range |
|---|---|---|
| 10-m wind component (m/s) | 0–36.36 | 0–31.52 |
| 10-m wind component (m/s) | 0–36.74 | 0–27.02 |
| Mean sea level pressure (Pa) | 91,303.00–106,822.00 | 93,162.31–105,149.50 |
| Geopotential @300hPa ( ) | 75,787.81–96,787.50 | 76,271.75–104,379.87 |
| Geopotential @500hPa ( ) | 43,307.90–59,109.62 | 43,784.96–63,338.88 |
| Target | Mean Dist. (km) | FPR | FNR | TPR | DR |
|---|---|---|---|---|---|
| 1 | 754.5759 | 0.0000 | 1.0000 | 0.0000 | 0 |
| 2 | 479.5773 | 0.0157 | 0.6000 | 0.4000 | 0 |
| 3 | 239.7887 | 0.0137 | 0.6000 | 0.4000 | 1 |
| 4 | 179.8415 | 0.0118 | 0.5000 | 0.5000 | 1 |
| Input | Reconstruction Error | Confidence (Anomaly Threshold) | ||
|---|---|---|---|---|
| 95% (1.156) | 98% (1.424) | 99% (6.838) | ||
| Clean | 0.825 | – | – | – |
| ALA | 164.884 | 100% | 100% | 100% |
| Cyc-Attack | 1.120 | 30% | 10% | 0% |
| Detector | 95% | 98% | 99% |
|---|---|---|---|
| Original PCA | 30% | 10% | 0% |
| Attack-Aware PCA | 80% | 70% | 0% |
Adversarial Attacks
Adversarial attacks exploit crafted inputs, poisoned data, or system interactions to manipulate AI predictions, behavior, privacy, or reliability.
See all 471 papers