生成對抗網路(Generative Adversarial Network, GAN)的技術不斷在進步，所產生的圖像單靠人眼已無法準確的判斷出其真偽性。在衛星圖資的分析與處理，GAN可以做為資料擴增的手段，但也有一些別有用心的應用(如資訊戰)，以GAN的手法造假圖片內容，混淆視聽，因此如何讓檢測模型分辨經過處理的衛星圖像之真偽性，是本論文的主要目標。
本論文從圖像處理的角度著手，將衛星圖像偽造偵測分成三個部分探討，第一部份為整張圖均為生成，研究三種不同衛星圖像（ProGAN、cGAN和CycleGAN）偽造方法，以及對圖像做強化處理是否會影響模型判別真偽圖的結果，並且添加對抗例攻擊能否有效讓模型偵測真假圖時混淆不清。第二部份是探討當衛星圖像中只有部份區域被經過修改時，模型是否能有效偵測被更動範圍。第三部份則探討不同的生成和偽造方式對於模型檢測效力的影響。
經實驗發現，不同GAN生成的衛星圖像可以很容易用機器學習的方式被辨別出來，但是當對圖像進行某些強化處理時，會讓檢測模型不容易分辨其真偽性，因此訓練了一個四分類的通用模型解決此問題。關於部份區域偽造實驗，發現測試在傳統相機拍出來的圖像上能夠正確偵測偽造區域，但是應用於衛星圖像時往往會失敗，推測可能是由於圖像類型的不同。關於衛星圖像切割成解析度小的圖像生成後再合併成原解析度之實驗，觀察到圖像在合併的地方會有很明顯的邊界存在，檢測模型也能很容易偵測出是合成的圖像，至於將真實圖像混合偽造圖片，由於偽造區塊只佔整張圖的一小部份，模型判斷此為真實的圖像。
統整本論文實驗結果，衛星圖資之真偽檢測，依其變造之手法有不同之難度，特別是部份區域偽造之偵測，仍是具挑戰之議題。

The technology of generative adversarial networks (GAN) is constantly evolving, and the synthesized images cannot be accurately distinguished by the human eyes alone. GAN has been applied to the analysis of satellite images, mostly for the purpose of data augmentation. Recently, however, we have seen a twist in its usage. In information warfare, GAN has been used to create fake satellite images or modify the image content by putting fake bridges, clouds to mislead or conceal important intelligence. To address the increasing counterfeit cases in satellite images, the goal of this thesis is to develop forgery detection algorithms that can classify fake images robustly and efficiently.
This thesis divides satellite image forgery detection into three parts. The first part deals with the case when the entire image is forged. Three satellite image synthesis methods, including ProGAN, cGAN and CycleGAN have been investigated. The effect of image pre-processing such as histogram equalization and bilateral filter has been evaluated. Adversarial attacks on the detection models have also been studied. The second part evaluates the effectiveness of the detection models when only certain parts of the satellite image have been modified. The third part discusses the impact of different generation and forgery methods on the detection model.
Experiments show that satellite images generated by different GANs can be easily identified by machine learning techniques. When enhancement is applied to the images, however, the detection model fails to distinguish its authenticity. A four-class classification model is proposed to address this issue. Regarding the experiments on partial forgery, it is found that methods which work for images taken by traditional cameras fail to perform satisfactorily on satellite images, possibly due to different image characteristics. Finally, when the input is partitioned into small areas and the individually GAN images are merged, discontinuities can be observed around the boundaries, making artifact the easy to detect. If we blend synthesized images with real images, the detection model shows difficulty when the size of the synthesized block is small.
The above experiments suggest that forgery detection in satellite images can be achieved when the whole image is synthesized. The task becomes more challenging if only part of the image is altered, an issue that deserves further investigation.

[1] Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., ... & Bengio, Y. (2014). Generative adversarial nets. Advances in neural information processing systems, 27.
[2] https://pansci.asia/archives/342421
[3] https://www.defenseone.com/technology/2019/03/next-phase-ai-deep-faking-whole-world-and-china-ahead/155944/
[4] https://medium.com/sentinel-hub/its-a-faaaake-or-not-bace4f0c01ec
[5] Zhao, B., Zhang, S., Xu, C., Sun, Y., & Deng, C. (2021). Deep fake geography? When geospatial data encounter Artificial Intelligence. Cartography and Geographic Information Science, 48(4), 338-352.
[6] Karras, T., Aila, T., Laine, S., & Lehtinen, J. (2017). Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196.
[7] https://blog.softwaremill.com/generative-adversarial-networks-in-satellite-image-datasets-augmentation-b7045d2f51ab
[8] Zhu, J. Y., Park, T., Isola, P., & Efros, A. A. (2017). Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision (pp. 2223-2232).
[9] Isola, P., Zhu, J. Y., Zhou, T., & Efros, A. A. (2017). Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 1125-1134).
[10] Wang, S. Y., Wang, O., Zhang, R., Owens, A., & Efros, A. A. (2020). Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 8695-8704).
[11] https://zhuanlan.zhihu.com/p/56165819
[12] Liang, J., Niu, L., & Zhang, L. (2021, July). Inharmonious region localization. In 2021 IEEE International Conference on Multimedia and Expo (ICME) (pp. 1-6). IEEE.
[13] Wu, Y., AbdAlmageed, W., & Natarajan, P. (2019). Mantra-net: Manipulation tracing network for detection and localization of image forgeries with anomalous features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 9543-9552).
[14] Zhang, X., Karaman, S., & Chang, S. F. (2019, December). Detecting and simulating artifacts in gan fake images. In 2019 IEEE International Workshop on Information Forensics and Security (WIFS) (pp. 1-6). IEEE.
[15] https://github.com/facebookresearch/pytorch_GAN_zoo
[16] https://github.com/softwaremill/sentinel-cgan
[17] https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix
[18] Goodfellow, I. J., Shlens, J., & Szegedy, C. (2014). Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572.
[19] Kurakin, A., Goodfellow, I., & Bengio, S. (2016). Adversarial examples in the physical world.
[20] Madry, A., Makelov, A., Schmidt, L., Tsipras, D., & Vladu, A. (2017). Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083.