Pub Date : 2026-01-12DOI: 10.1109/LSP.2026.3653395
Luan Portella;R. J. Cintra;Aluísio Pinheiro
This letter introduces low-complexity approximations for the Wavelet Energy Correlation Screening (WECS) method, which aims at change detection in multitemporal SAR images. The WECS method relies on the non-decimated discrete wavelet transform (ND-DWT) to compute approximation coefficients employed in a feature screening process based on the Pearson correlation. Although effective, WECS presents a high computational cost due to its repeated wavelet filtering stage. To overcome this drawback, we propose two approximations for the wavelet filter coefficients, obtained by truncating their canonical signed digit (CSD) representation, which significantly reduces the number of arithmetic operations. Numerical experiments using both simulated and real-world datasets demonstrate that the proposed methods not only maintain the performance of the original WECS but also achieve computational gains, even outperforming it in certain scenarios.
{"title":"Low-Complexity Approximations of the WECS Method for SAR Change Detection","authors":"Luan Portella;R. J. Cintra;Aluísio Pinheiro","doi":"10.1109/LSP.2026.3653395","DOIUrl":"https://doi.org/10.1109/LSP.2026.3653395","url":null,"abstract":"This letter introduces low-complexity approximations for the Wavelet Energy Correlation Screening (WECS) method, which aims at change detection in multitemporal SAR images. The WECS method relies on the non-decimated discrete wavelet transform (ND-DWT) to compute approximation coefficients employed in a feature screening process based on the Pearson correlation. Although effective, WECS presents a high computational cost due to its repeated wavelet filtering stage. To overcome this drawback, we propose two approximations for the wavelet filter coefficients, obtained by truncating their canonical signed digit (CSD) representation, which significantly reduces the number of arithmetic operations. Numerical experiments using both simulated and real-world datasets demonstrate that the proposed methods not only maintain the performance of the original WECS but also achieve computational gains, even outperforming it in certain scenarios.","PeriodicalId":13154,"journal":{"name":"IEEE Signal Processing Letters","volume":"33 ","pages":"643-647"},"PeriodicalIF":3.9,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081993","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1109/LSP.2026.3652955
Min Li;Zhaofei Hao;Gang Li;Jin Wan;Delong Han;Mingle Zhou
Small Object Detection (SOD) aims to accurately identify and locate small objects in images. However, existing methods usually focus on exploring spatial domain features, neglecting high-frequency features that preserve fine-grained details such as texture and edge information. To overcome this limitation, we propose a High-Frequency Feature-Oriented Network (HFFO-Net). First, we introduce the Channel-wise Frequency Modulation Module (CFMM), which leverages the 2D Discrete Cosine Transform (DCT) to accentuate salient frequency components while mitigating noise interference. Second, we design a High-Frequency Oriented Module (HFOM), which utilizes the Channel Selection Branch (CSB) and Spatial Selection Branch (SSB) to highlight small objects in the channel and spatial region. Third, we introduce a Dual-Query Attention Fusion Mechanism (DQAFM), which reduces the semantic gap between spatial and frequency features and achieves better feature fusion through bidirectional cross-attention. Extensive experiments are implemented, and the corresponding results demonstrate that HFFO-Net excels at detecting small objects.
{"title":"Boosting Small Object Detection via High-Frequency Feature Oriented Network","authors":"Min Li;Zhaofei Hao;Gang Li;Jin Wan;Delong Han;Mingle Zhou","doi":"10.1109/LSP.2026.3652955","DOIUrl":"https://doi.org/10.1109/LSP.2026.3652955","url":null,"abstract":"Small Object Detection (SOD) aims to accurately identify and locate small objects in images. However, existing methods usually focus on exploring spatial domain features, neglecting high-frequency features that preserve fine-grained details such as texture and edge information. To overcome this limitation, we propose a High-Frequency Feature-Oriented Network (HFFO-Net). First, we introduce the Channel-wise Frequency Modulation Module (CFMM), which leverages the 2D Discrete Cosine Transform (DCT) to accentuate salient frequency components while mitigating noise interference. Second, we design a High-Frequency Oriented Module (HFOM), which utilizes the Channel Selection Branch (CSB) and Spatial Selection Branch (SSB) to highlight small objects in the channel and spatial region. Third, we introduce a Dual-Query Attention Fusion Mechanism (DQAFM), which reduces the semantic gap between spatial and frequency features and achieves better feature fusion through bidirectional cross-attention. Extensive experiments are implemented, and the corresponding results demonstrate that HFFO-Net excels at detecting small objects.","PeriodicalId":13154,"journal":{"name":"IEEE Signal Processing Letters","volume":"33 ","pages":"584-588"},"PeriodicalIF":3.9,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026329","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1109/LSP.2026.3652127
Kun Lu;Hongli Zhang;Yuchen Yang;Chao Meng;Binxing Fang
Social bots constitute a substantial fraction of active accounts on digital platforms, fundamentally threatening information authenticity and democratic discourse. Contemporary detection methods confront critical limitations: information imbalance across heterogeneous relations, computational challenges in processing massive neighborhoods, and inadequate multi-scale representation learning. We propose HIER (Heterogeneous Information Bottleneck and Expert Routing), a pioneering framework that integrates variational information theory with mixture-of-experts paradigms for social network analysis. HIER introduces relation-aware variational information bottleneck for optimal compression across relationship types, dynamic sparse expert routing that extends mixture-of-experts to edge-level graph processing, and dual-scale mutual information maximization enhancing representation discriminability through neighborhood consistency and graph-level contrastive learning. Experimental validation demonstrates HIER’s superior performance across real-world datasets, establishing new benchmarks for heterogeneous social bot detection.
{"title":"HIER: Heterogeneous Information Bottleneck and Expert Routing for Social Bot Detection","authors":"Kun Lu;Hongli Zhang;Yuchen Yang;Chao Meng;Binxing Fang","doi":"10.1109/LSP.2026.3652127","DOIUrl":"https://doi.org/10.1109/LSP.2026.3652127","url":null,"abstract":"Social bots constitute a substantial fraction of active accounts on digital platforms, fundamentally threatening information authenticity and democratic discourse. Contemporary detection methods confront critical limitations: information imbalance across heterogeneous relations, computational challenges in processing massive neighborhoods, and inadequate multi-scale representation learning. We propose HIER (Heterogeneous Information Bottleneck and Expert Routing), a pioneering framework that integrates variational information theory with mixture-of-experts paradigms for social network analysis. HIER introduces relation-aware variational information bottleneck for optimal compression across relationship types, dynamic sparse expert routing that extends mixture-of-experts to edge-level graph processing, and dual-scale mutual information maximization enhancing representation discriminability through neighborhood consistency and graph-level contrastive learning. Experimental validation demonstrates HIER’s superior performance across real-world datasets, establishing new benchmarks for heterogeneous social bot detection.","PeriodicalId":13154,"journal":{"name":"IEEE Signal Processing Letters","volume":"33 ","pages":"521-525"},"PeriodicalIF":3.9,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026347","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1109/LSP.2026.3652954
Kai-Wei Peng
We study sampling of smooth/bandlimited graph signals when (i) sensor noise is heterogeneous across vertices and (ii) the graph used to design the sampler can be mildly mismatched to the true topology.We propose a risk-aware variant of local low-pass importance sampling that scores each vertex via a Hutchinson estimator of the diagonal of a graph heat-kernel operator and reweights the score by the inverse noise variance. The sampler selects without replacement according to these risk-aware scores. Reconstruction is performed with standard decoders (Tikhonov, Bandlimited, and a Chebyshev data-consistent smoother), enabling fair comparisons to prior work. On grid, Erdős–Rényi (ER), and Barabási–Albert (BA) graphs, our approach consistently reduces the normalized root-mean-square error (NRMSE) compared to random sampling; the gain increases with the sampling rate and persists under selection-graph mismatch. The method is simple, eigendecomposition-free, and scales linearly in the number of edges per Hutchinson probe.
{"title":"Risk-Aware Low-Pass Importance Sampling for Graph Signals Under Heterogeneous Noise and Model Mismatch","authors":"Kai-Wei Peng","doi":"10.1109/LSP.2026.3652954","DOIUrl":"https://doi.org/10.1109/LSP.2026.3652954","url":null,"abstract":"We study sampling of smooth/bandlimited graph signals when (i) sensor noise is heterogeneous across vertices and (ii) the graph used to design the sampler can be mildly mismatched to the true topology.We propose a risk-aware variant of local low-pass importance sampling that scores each vertex via a Hutchinson estimator of the diagonal of a graph heat-kernel operator and reweights the score by the inverse noise variance. The sampler selects without replacement according to these risk-aware scores. Reconstruction is performed with standard decoders (<sc>Tikhonov</small>, <sc>Bandlimited</small>, and a Chebyshev data-consistent smoother), enabling fair comparisons to prior work. On grid, Erdős–Rényi (ER), and Barabási–Albert (BA) graphs, our approach consistently reduces the normalized root-mean-square error (NRMSE) compared to random sampling; the gain increases with the sampling rate and persists under selection-graph mismatch. The method is simple, eigendecomposition-free, and scales linearly in the number of edges per Hutchinson probe.","PeriodicalId":13154,"journal":{"name":"IEEE Signal Processing Letters","volume":"33 ","pages":"556-558"},"PeriodicalIF":3.9,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026402","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1109/LSP.2026.3653238
Da-Hee Yang;Joon-Hyuk Chang
Noise-robust automatic speech recognition (ASR) has been commonly addressed by applying speech enhancement (SE) at the waveform level before recognition. However, speech-level enhancement does not always translate into consistent recognition improvements due to residual distortions and mismatches with the latent space of the ASR encoder. In this letter, we introduce a complementary strategy termed latent-level enhancement, where distorted representations are refined during ASR inference. Specifically, we propose a plug-and-play Flow Matching Refinement module (FM-Refiner) that operates on the output latents of a pretrained CTC-based ASR encoder. Trained to map imperfect latents—either directly from noisy inputs or from enhanced-but-imperfect speech—toward their clean counterparts, the FM-Refiner is applied only at inference, without fine-tuning ASR parameters. Experiments show that FM-Refiner consistently reduces word error rate, both when directly applied to noisy inputs and when combined with conventional SE front-ends. These results demonstrate that latent-level refinement via flow matching provides a lightweight and effective complement to existing SE approaches for robust ASR.
{"title":"Latent-Level Enhancement With Flow Matching for Robust Automatic Speech Recognition","authors":"Da-Hee Yang;Joon-Hyuk Chang","doi":"10.1109/LSP.2026.3653238","DOIUrl":"https://doi.org/10.1109/LSP.2026.3653238","url":null,"abstract":"Noise-robust automatic speech recognition (ASR) has been commonly addressed by applying speech enhancement (SE) at the waveform level before recognition. However, speech-level enhancement does not always translate into consistent recognition improvements due to residual distortions and mismatches with the latent space of the ASR encoder. In this letter, we introduce a complementary strategy termed latent-level enhancement, where distorted representations are refined during ASR inference. Specifically, we propose a plug-and-play Flow Matching Refinement module (FM-Refiner) that operates on the output latents of a pretrained CTC-based ASR encoder. Trained to map imperfect latents—either directly from noisy inputs or from enhanced-but-imperfect speech—toward their clean counterparts, the FM-Refiner is applied only at inference, without fine-tuning ASR parameters. Experiments show that FM-Refiner consistently reduces word error rate, both when directly applied to noisy inputs and when combined with conventional SE front-ends. These results demonstrate that latent-level refinement via flow matching provides a lightweight and effective complement to existing SE approaches for robust ASR.","PeriodicalId":13154,"journal":{"name":"IEEE Signal Processing Letters","volume":"33 ","pages":"589-593"},"PeriodicalIF":3.9,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026484","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1109/LSP.2026.3652951
Weicheng Gao;Shui Liu;Jinshuo Wang;Xiaodong Qu;Xiaopeng Yang
Through-the-wall radar (TWR) can monitor and analyze the motion characteristics and activity patterns of indoor human targets, with the advantages of non-contact, high flexibility and privacy protection. However, existing TWR human activity recognition (HAR) techniques developed based on single-channel radar contain limited Doppler information, making it difficult to achieve accurate recognition on data where the direction of human motion is not parallel to the radar observation. To solve this problem, in this letter, a multi-input-multi-output (MIMO) TWR micro-Doppler signature augmentation method based on multi-channel information fusion is proposed. First, a multi-channel Doppler profile feature fusion method based on multi-scale wavelets with low-rank decomposition is presented. Then, a motion parameter estimation method based on Broyden-Fletcher-Goldfarb-Shanno (BFGS) global optimization is proposed, and the fused Doppler profile transformation is implemented using the obtained orientation of human motion. Numerical simulated and measured experiments demonstrate the effectiveness of the proposed method.
{"title":"MIMO Through-the-Wall Radar Micro-Doppler Signature Augmentation Method Based on Multi-Channel Information Fusion","authors":"Weicheng Gao;Shui Liu;Jinshuo Wang;Xiaodong Qu;Xiaopeng Yang","doi":"10.1109/LSP.2026.3652951","DOIUrl":"https://doi.org/10.1109/LSP.2026.3652951","url":null,"abstract":"Through-the-wall radar (TWR) can monitor and analyze the motion characteristics and activity patterns of indoor human targets, with the advantages of non-contact, high flexibility and privacy protection. However, existing TWR human activity recognition (HAR) techniques developed based on single-channel radar contain limited Doppler information, making it difficult to achieve accurate recognition on data where the direction of human motion is not parallel to the radar observation. To solve this problem, in this letter, a multi-input-multi-output (MIMO) TWR micro-Doppler signature augmentation method based on multi-channel information fusion is proposed. First, a multi-channel Doppler profile feature fusion method based on multi-scale wavelets with low-rank decomposition is presented. Then, a motion parameter estimation method based on Broyden-Fletcher-Goldfarb-Shanno (BFGS) global optimization is proposed, and the fused Doppler profile transformation is implemented using the obtained orientation of human motion. Numerical simulated and measured experiments demonstrate the effectiveness of the proposed method.","PeriodicalId":13154,"journal":{"name":"IEEE Signal Processing Letters","volume":"33 ","pages":"579-583"},"PeriodicalIF":3.9,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026499","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1109/LSP.2026.3652122
Bingzhang Wu;Shaoxuan Li;Ziyao Pan;Rongxin Zhang;Wei Su
This letter proposes a lightweight parallel recurrent–convolutional scheme to improve generalization capability and recognition accuracy while maintaining low computational complexity in resource-constrained underwater acoustic channels. In this scheme, the lightweight convolutional network is used to extract time–frequency features, and the lightweight recurrent network with gated recurrent units is used to capture long-term temporal phase correlations, thereby alleviating the Doppler-induced phase rotation and inter-symbol interference in time-varying multipath underwater acoustic channels. Sea-trial data are collected during shallow-water sea trials with strictly separated training and evaluation datasets. Experimental results on ten underwater acoustic modulation types show that the proposed scheme improves recognition accuracy by 6.2% and reduces computational cost by 22.4%, while exhibiting stronger generalization capability compared with benchmark schemes.
{"title":"LRCPN: A Lightweight Parallel Scheme for Underwater Acoustic Modulation Recognition","authors":"Bingzhang Wu;Shaoxuan Li;Ziyao Pan;Rongxin Zhang;Wei Su","doi":"10.1109/LSP.2026.3652122","DOIUrl":"https://doi.org/10.1109/LSP.2026.3652122","url":null,"abstract":"This letter proposes a lightweight parallel recurrent–convolutional scheme to improve generalization capability and recognition accuracy while maintaining low computational complexity in resource-constrained underwater acoustic channels. In this scheme, the lightweight convolutional network is used to extract time–frequency features, and the lightweight recurrent network with gated recurrent units is used to capture long-term temporal phase correlations, thereby alleviating the Doppler-induced phase rotation and inter-symbol interference in time-varying multipath underwater acoustic channels. Sea-trial data are collected during shallow-water sea trials with strictly separated training and evaluation datasets. Experimental results on ten underwater acoustic modulation types show that the proposed scheme improves recognition accuracy by 6.2% and reduces computational cost by 22.4%, while exhibiting stronger generalization capability compared with benchmark schemes.","PeriodicalId":13154,"journal":{"name":"IEEE Signal Processing Letters","volume":"33 ","pages":"516-520"},"PeriodicalIF":3.9,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026512","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1109/LSP.2026.3653403
Lu Li;Qinkun Xiao;Peiran Liu
Continuous sign language recognition (CSLR) requires fine-grained alignment between visual sequences and gloss annotations under weak supervision, which is challenged by modality heterogeneity and ambiguous frame-to-gloss correspondence. We propose a Multimodal Cosine Similarity Transformer (MMCST) to address these issues. MMCST integrates RGB and keypoint heatmap features via gated fusion, and aligns them with gloss embeddings through a Gloss-Conditioned Cosine-Normalized Attention (GCNA) mechanism that stabilizes cross-modal alignment. To further enhance semantic consistency, we introduce Gloss-aware Contrastive Regularization (GLCR). The fused representation is modeled by a cosine-similarity Transformer and decoded with CTC. Experimental results show that MMCST achieves consistent improvements over strong baselines, and ablation studies confirm the effectiveness of gated fusion, GCNA, and GLCR in improving semantic alignment and yielding smoother training dynamics.
{"title":"Multimodal Cosine Similarity Transformer for Gloss-Guided Sign Language Recognition","authors":"Lu Li;Qinkun Xiao;Peiran Liu","doi":"10.1109/LSP.2026.3653403","DOIUrl":"https://doi.org/10.1109/LSP.2026.3653403","url":null,"abstract":"Continuous sign language recognition (CSLR) requires fine-grained alignment between visual sequences and gloss annotations under weak supervision, which is challenged by modality heterogeneity and ambiguous frame-to-gloss correspondence. We propose a Multimodal Cosine Similarity Transformer (MMCST) to address these issues. MMCST integrates RGB and keypoint heatmap features via gated fusion, and aligns them with gloss embeddings through a Gloss-Conditioned Cosine-Normalized Attention (GCNA) mechanism that stabilizes cross-modal alignment. To further enhance semantic consistency, we introduce Gloss-aware Contrastive Regularization (GLCR). The fused representation is modeled by a cosine-similarity Transformer and decoded with CTC. Experimental results show that MMCST achieves consistent improvements over strong baselines, and ablation studies confirm the effectiveness of gated fusion, GCNA, and GLCR in improving semantic alignment and yielding smoother training dynamics.","PeriodicalId":13154,"journal":{"name":"IEEE Signal Processing Letters","volume":"33 ","pages":"673-677"},"PeriodicalIF":3.9,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146082110","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-12DOI: 10.1109/LSP.2026.3652119
Bo Tang;Steven Kay;Kaushallya Adhikari
A classification rule based on the cumulant generating function of the training data, called the Cumulant Generating Function Classifier (CGFC), has been recently proposed, and has shown promising performance in terms of improved classification accuracy and robustness against noises. This paper first presents a new information-theoretical explanation of CGFC which indeed makes a classification by minimizing sample mutual information. The original CGFC is a type of global model, and a new variant, called Local-CGFC, is further introduced in this paper to achieve a local classification rule. Experimental studies on real-life datasets demonstrate the effectiveness of the proposed classifier and further illustrate its great potential for a number of real-world applications.
{"title":"Local-CGFC: A Local Cumulant Generating Function Classification Rule","authors":"Bo Tang;Steven Kay;Kaushallya Adhikari","doi":"10.1109/LSP.2026.3652119","DOIUrl":"https://doi.org/10.1109/LSP.2026.3652119","url":null,"abstract":"A classification rule based on the cumulant generating function of the training data, called the Cumulant Generating Function Classifier (CGFC), has been recently proposed, and has shown promising performance in terms of improved classification accuracy and robustness against noises. This paper first presents a new information-theoretical explanation of CGFC which indeed makes a classification by minimizing sample mutual information. The original CGFC is a type of global model, and a new variant, called Local-CGFC, is further introduced in this paper to achieve a local classification rule. Experimental studies on real-life datasets demonstrate the effectiveness of the proposed classifier and further illustrate its great potential for a number of real-world applications.","PeriodicalId":13154,"journal":{"name":"IEEE Signal Processing Letters","volume":"33 ","pages":"546-550"},"PeriodicalIF":3.9,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026463","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-01-06DOI: 10.1109/LSP.2026.3651083
Guobiao Li;Sheng Li;Zhenxing Qian;Xinpeng Zhang
Deep functionality hiding is an emerging technique that embeds confidential or sensitive functions within seemingly benign deep learning models (DLMs), which perform ordinary machine learning tasks. This enables such models to execute covert tasks while remaining undetected. Despite the rapid progress in deep functionality hiding, countermeasures remain unexplored. In this paper, we propose Distribution Offset Analysis (DOA), a novel method for detecting hidden functionalities in DLMs. Our key insight is that the weight distribution of a benign DLM typically follows a Gaussian distribution, whereas a container DLM with hidden functionalities exhibits notable statistical deviations from this Gaussian pattern. In our methodology, we first compute the distributional distance (i.e., offsets) between the model's weights and an ideal Gaussian distribution. We then fuse these offsets with weight features into a unified representation, which is subsequently used to train a meta-classifier for hidden functionality detection. Through extensive experiments, we demonstrate the effectiveness of the proposed DOA method, which achieves an average detection rate of over 87% against existing state-of-the-art deep functionality hiding techniques.
{"title":"Toward Detecting Hidden Functionalities in Deep Learning Models","authors":"Guobiao Li;Sheng Li;Zhenxing Qian;Xinpeng Zhang","doi":"10.1109/LSP.2026.3651083","DOIUrl":"https://doi.org/10.1109/LSP.2026.3651083","url":null,"abstract":"Deep functionality hiding is an emerging technique that embeds confidential or sensitive functions within seemingly benign deep learning models (DLMs), which perform ordinary machine learning tasks. This enables such models to execute covert tasks while remaining undetected. Despite the rapid progress in deep functionality hiding, countermeasures remain unexplored. In this paper, we propose Distribution Offset Analysis (DOA), a novel method for detecting hidden functionalities in DLMs. Our key insight is that the weight distribution of a benign DLM typically follows a Gaussian distribution, whereas a container DLM with hidden functionalities exhibits notable statistical deviations from this Gaussian pattern. In our methodology, we first compute the distributional distance (<italic>i.e.,</i> offsets) between the model's weights and an ideal Gaussian distribution. We then fuse these offsets with weight features into a unified representation, which is subsequently used to train a meta-classifier for hidden functionality detection. Through extensive experiments, we demonstrate the effectiveness of the proposed DOA method, which achieves an average detection rate of over 87% against existing state-of-the-art deep functionality hiding techniques.","PeriodicalId":13154,"journal":{"name":"IEEE Signal Processing Letters","volume":"33 ","pages":"541-545"},"PeriodicalIF":3.9,"publicationDate":"2026-01-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146026418","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：