Adversarial Attacks as a Diagnostic Tool for a Mining Detection Network in the Brazilian Legal Amazon

• Leonardo Fajardo Grupioni USP
• Gabriel Stephano Santos USP
• Paulo Sérgio Cugnasca USP
• Felipe Valencia de Almeida USP

DOI: https://doi.org/10.5753/sbseg_estendido.2026.33861

Resumo

Convolutional neural networks (CNNs) applied to satellite imagery are increasingly adopted as the standard tool to detect mining in the Brazilian Legal Amazon, as their predicition start to feed environmental enforcement pipelines. We trained such a CNN and obtained an F1-Score of 0.9247 with an AUC of 0.9930. Adversarial attacks show that these numbers hide what the model learned. With the addition of random noise in the data, up to 45% of the of forest patches were missclassified as mining, while low-pass noise with the same budget and the same mean shift turns 0.5%, so the decision rests on high frequency texture. The fragility is asymmetric, since one FGSM step at 4/255 converts every forest patch into mining but evades only 5% of the real mining.

Referências

Camalan, S., Cui, K., Pauca, V. P., Alqahtani, S., Silman, M., Chan, R., Plemmons, R. J., Dethier, E. N., Fernandez, L. E., and Lutz, D. A. (2022). Change detection of Amazonian alluvial gold mining using deep learning and Sentinel-2 imagery. Remote Sensing, 14(7):1746.

Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F. A., and Brendel, W. (2019). ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations (ICLR).

Global Initiative Against Transnational Organized Crime (2023). Amazon underworld: economias criminosas na maior floresta tropical do mundo. Technical report, GI-TOC and Amazon Watch and InfoAmazonia, Geneva, Switzerland. Institutional report.

Goodfellow, I. J., Shlens, J., and Szegedy, C. (2015). Explaining and harnessing adversarial examples. In International Conference on Learning Representations (ICLR).

Grupioni, L., Gallois, T. J., and Almeida, F. (2026). A sentinel-2 image dataset for mining detection across mining proportion ranges in the brazilian legal amazon. In Anais do XVII Workshop de Computação Aplicada à Gestão do Meio Ambiente e Recursos Naturais, pages 348–351, Porto Alegre, RS, Brasil. SBC.

Lemes Neto, N., Galo, M. d. L. B. T., de Oliveira, R. A., Watanabe, F. S. Y., and Galo, M. (2026). Towards SAR-Based monitoring of illegal mining in the Brazilian Amazon using convolutional neural networks. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, X-3-W4-2025:205–212.

Xu, Y. and Ghamisi, P. (2022). Universal adversarial examples in remote sensing: Methodology and benchmark. IEEE Transactions on Geoscience and Remote Sensing, 60:1–15.