Amplitude SAR Imagery Splicing Localization

Edoardo Daniele Cannas, Nicolò Bonettini, Sara Mandelli, Paolo Bestagini, Stefano TubaroarXiv 2022 · DOI · arXiv

TL;DR

This paper introduces the first dedicated framework for localizing splicing forgeries (copy-paste manipulations) in amplitude Synthetic Aperture Radar (SAR) imagery. By adapting deep learning-based noise residual extraction to the unique processing chain of radar data, the proposed method successfully flags tampered regions, significantly outperforming forensic tools designed for standard optical photographs.

What Problem Does This Paper Address?

With the rise of open-access satellite data, Synthetic Aperture Radar (SAR) imagery has become vital for environmental monitoring, mapping, and security. However, like standard digital photos, SAR images are vulnerable to malicious editing, such as splicing—where a region from one image is pasted into another to hide or fabricate objects on the ground.

While a vast library of multimedia forensic tools exists to detect optical image tampering, they fail when applied to SAR data. Optical images leave distinct footprints related to camera sensors (like color filter array interpolation). SAR images, conversely, are formed by active radar sensors on a moving platform processing backscattered echoes. This raw data undergoes a complex pipeline of focusing, geocoding, sub-swath merging, and product slicing, leaving a completely different set of statistical artifacts. This paper fills the gap by developing a forensic method explicitly tailored to the unique lifecycle of amplitude SAR images.

Methods

The authors propose a two-stage pipeline to detect and localize spliced regions:

Proposed processing pipeline
  1. Fingerprint Extraction: The system uses a modified Denoising Convolutional Neural Network (DnCNN) to suppress the high-level semantic content of the image and isolate a "fingerprint" of the underlying processing traces. While optical forensic tools (like Noiseprint) look for spatially periodic camera artifacts, SAR artifacts are product-wide. The authors relax the spatial constraints during training, forcing the network to learn features common to an entire SAR product rather than a specific pixel grid location. They test three versions, finding the best results with the Augmented SAR Adapted Extractor (ASAE), which uses resizing during training to mimic different SAR processing grids.
  2. Tampering Mask Estimation: To convert the continuous fingerprint into a binary mask (tampered vs. pristine), the authors evaluate three segmentation strategies:
  3. Unsupervised K-means: Clusters fingerprint patches based on Euclidean distance and identifies the most compact cluster as the forgery.
  4. Unsupervised Gaussian Mixture Models (GMM): Uses expectation-maximization to statistically isolate the anomalous spliced region.
  5. Supervised U-Net: A deep learning segmentation network trained on known tampered fingerprints to directly predict the binary mask.

Datasets

The authors constructed custom datasets using 20 high-resolution Ground Range Detected (GRD) products from the Copernicus Sentinel-1B satellite mission, sliced into $1024 \times 1024$ pixel tiles:

  • Fingerprint Extraction Dataset (FED): 4,000 pristine tiles used to train the baseline fingerprint extractor.
  • Spliced Dataset 1 (SD1): 1,600 inter-spliced tiles (where donor and target regions come from different SAR products) used to train the supervised U-Net.
  • Spliced Dataset 2 (SD2): A challenging testing dataset of 7,000 tiles. It is split equally between inter-splicing and intra-splicing (where the donor patch comes from the same SAR product). It includes realistic attacker manipulations like Gaussian/Laplacian noise, blurring, and speckle noise.

Key Insights and Ideas

  • SAR Products as "Devices": In optical forensics, networks are trained to differentiate between camera models. This paper re-frames the problem for satellite imagery: the network learns to differentiate between individual SAR products, treating their distinct processing histories as unique "device" signatures.
  • Relaxing Spatial Constraints: Dropping the rigid grid-alignment constraints used in optical forensics is crucial. Because SAR data is concatenated and resampled globally during product generation, focusing on product-wide consistency yields a far more effective fingerprint.
  • Data Augmentation as Synthetic History: By resizing tiles during training and treating them as entirely new SAR products, the ASAE model becomes highly sensitive to resampling operations. This allows it to generalize remarkably well to completely unseen post-processing edits like blurring or noise addition.

Key Results

The proposed ASAE framework, paired with either GMM or U-Net mask estimation, consistently outperformed state-of-the-art optical forensic methods like Noiseprint and Splicebuster.

  • Inter-Splicing Detection: In scenarios where the pasted patch came from a different satellite product, the ASAE method achieved an Intersection Over Union (IOU) greater than 0.66 and a Balanced Accuracy (BA) exceeding 0.91 across almost all editing types (such as rotation, resizing, and blurring).
  • Robustness to Blurring: For average blurring attacks, the ASAE-GMM pipeline achieved an IOU of 0.859, compared to Noiseprint's 0.748 and Splicebuster's 0.478.
  • Intra-Splicing Performance: Splicing within the same product is significantly harder to detect because the background and foreground share the same processing history. However, if the attacker performs any scaling or blurring to make the patch fit visually, the pipeline successfully localizes the forgery, with the U-Net achieving a BA of up to 0.95.

Limitations and Future Work

The primary blindspot of the proposed method is "No Editing" intra-splicing. If an attacker copies a region and pastes it elsewhere within the exact same SAR product without resizing, rotating, or adding noise, the pipeline cannot detect it (yielding a BA of ~0.50, equivalent to random guessing). Because the donor and target share an identical processing chain and no modifications were made, no statistical inconsistencies are introduced into the fingerprint.

Future work aims to address this limitation by exploring more complex splicing attacks (such as GAN-generated inpainting), incorporating physics-based radar properties, and adapting these localization pipelines to electro-optical satellite imagery.

Summary updated

Abstract

Synthetic Aperture Radar (SAR) images are a valuable asset for a wide variety of tasks. In the last few years, many websites have been offering them for free in the form of easy to manage products, favoring their widespread diffusion and research work in the SAR field. The drawback of these opportunities is that such images might be exposed to forgeries and manipulations by malicious users, raising new concerns about their integrity and trustworthiness. Up to now, the multimedia forensics literature has proposed various techniques to localize manipulations in natural photographs, but the integrity assessment of SAR images was never investigated. This task poses new challenges, since SAR images are generated with a processing chain completely different from that of natural photographs. This implies that many forensics methods developed for natural images are not guaranteed to succeed. In this paper, we investigate the problem of amplitude SAR imagery splicing localization. Our goal is to localize regions of an amplitude SAR image that have been copied and pasted from another image, possibly undergoing some kind of editing in the process. To do so, we leverage a Convolutional Neural Network (CNN) to extract a fingerprint highlighting inconsistencies in the processing traces of the analyzed input. Then, we examine this fingerprint to produce a binary tampering mask indicating the pixel region under splicing attack. Results show that our proposed method, tailored to the nature of SAR signals, provides better performances than state-of-the-art forensic tools developed for natural images.
BibTeX
@misc{cannas2022amplitude,
  author = {Edoardo Daniele Cannas and Nicolò Bonettini and Sara Mandelli and Paolo Bestagini and Stefano Tubaro},
  title = {{Amplitude SAR Imagery Splicing Localization}},
  year = {2022},
  eprint = {2201.02409},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2201.02409},
}
Figure