19 June 2025 in Proceedings of the ACM on Software Engineering
Learning-based Android malware detection has earned significant recognition across industry and academia, yet its effectiveness hinges on the accuracy of labeled training data. Manual labeling, being prohibitively expensive, has prompted the use of automated methods, such as leveraging anti-virus engines like VirusTotal, which unfortunately introduces mislabeling, aka "label noise". The state-of-the-art label noise reduction approach, MalWhiteout, can effectively reduce random label noise but underperforms in mitigating real-world emergent malware (EM) label noise stemming from newly emerging Android malware variants overlooked by VirusTotal. To tackle this, we conceptualize EM label noise detection as an anomaly detection problem and introduce a novel tool, MalCleanse, that surpasses MalWhiteout's ability to address EM label noise. MalCleanse combines uncertainty estimation with unsupervised anomaly detection, identifying samples with high uncertainty as mislabeled, thereby enhancing its capability to remove EM label noise. Our experimental results demonstrate a significant reduction in EM label noise by approximately 25.25%, achieving an F1 Score of 80.32% for label noise detection at a noise ratio of 40.61%. Notably, MalCleanse outperforms MalWhiteout with an increase of 40.9% in overall F1 score for mitigating EM label noise. This paper pioneers the integration of deep neural network model uncertainty to refine label accuracy, thereby enhancing the reliability of malware detection systems. Our approach represents a significant step forward in addressing the challenges posed by emergent malware in automated labeling systems.
28 January 2025 in IEEE Geoscience and Remote Sensing Letters
Feature representation plays a key role in matching keypoints, especially for the multispectral images of large spectral difference. On such image pairs, existing methods typically use the two images only, but it is challenging to directly learn spectrum invariant feature representation due to the complex non-linear distortion between them. To address this issue, this paper proposes using intermediate-band images to facilitate learning spectrum invariant feature representation. For this purpose, this work designs a spectrum adaptive gate network (SAG-Net) that consists of a spectrum gate (SPeG) module and a deep feature extractor. The SPeG module selectively activates the spectrum invariant features according to input image content on-the-fly. It hence allows for training on the images of over two bands simultaneously with a single network without the need of an individual branch per band. In order to investigate the SPeG module, we also constructed a L9-MSI dataset including 3167 scenes of aligned images across five spectral bands (visible, B5, B6, B7, and B10) from the Landsat 9 imagery. The experimental results demonstrate the SPeG module can learn the common feature representation for varying-band images, and the intermediate B5, B6, and B7 images are useful for the SAG-Net to learn the common feature between the visible and B10. On the L9-MSI dataset, the SAG-Net significantly improved the number of correct matches and the matching score. Our dataset will be released at https://github.com/bohanlee/L9MSI-Dataset.git.
5 June 2025 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
This paper proposes a method composed of a loss function and a feature extractor structure, to learn the distinctive feature representation (DIFR) for descriptors on multimodal images. The distinctiveness of the feature representation plays a key role in matching keypoints, and has drawn a lot of research attention. In the past, many methods were proposed that either employed handcrafted features or deep neural networks to learn the feature representation. However, the incorrect ratio of built keypoint matches still remains high due to the insufficient distinctiveness of the associated descriptors. To address this problem, this work designs a new contrastive loss function that for the first time incorporates multiple negative descriptors from both the reference image and the test image, to constrain the networks to learn the distinctive feature representation for descriptors. To further enable the descriptors to encode more global information, we design RestExtractor, a novel CNN-Transformer structure that effectively adapts the self-attention mechanism to the feature learning for multimodal images. In the RestExtractor, we design a gated-v block and a receptive field enhanced feed-forward network for extracting distinctive modal-shared feature representation. We evaluated the proposed method on two challenging multimodal image datasets, VIS-SAR and VIS-IR. The experimental results showed that the proposed DIFR significantly improved the number of correct keypoint matches over the state-of-the-art methods. Our source code will be released on https://github.com/bohanlee/DIFR/
1 January 2026 in IEEE Transactions on Cognitive Communications and Networking