The integrity analysis of compressed videos is becoming increasingly important in modern visual communication systems. In practical workflows, manipulated or republished videos are often decoded and re-encoded during editing, transcoding, or delivery adaptation, making Double Compression Detection (DCD) a fundamental problem in digital video forensics. This problem becomes more challenging under Versatile Video Coding (VVC), a representative next-generation video coding standard, because its advanced prediction tools complicate codec-side recompression traces. Existing DCD methods usually rely on conventional spatio-temporal representations or stage-wise multi-round compression statistics, making it difficult to jointly exploit decision-level transition evidence and stage-dependent discrepancy cues across successive compressions in VVC. To address this issue, we propose a Compressional-Temporal Prediction Decision Evolution Network (CTPDE-Net), which jointly exploits decision-level transition evidence and stage-dependent discrepancy cues across successive compressions. Within this framework, Prediction Decision Co-occurrence Mapping (PDCM) organizes spatially aligned transitions of intra prediction tools, while a plug-and-play Cross-Compression Discrepancy Amplifier module (CCDA-module) adaptively enhances informative discrepancy cues to capture discriminative recompression-induced changes. Experimental results demonstrate that, under the evaluated VVC recompression settings, the proposed method achieves the best overall performance among the compared methods and maintains consistent effectiveness across the evaluated backbone architectures.

The proliferation of highly realistic synthetic faces poses significant challenges to information security, necessitating robust and generalizable Deepfake detection systems. Existing multi-stream networks rely on feature-level fusion to integrate complementary cues but often overlook the critical importance of data quality. In this work, we propose Iterative Consensus Learning (ICL), a novel framework that introduces data-level multi-stream assessment via collaborative sample evaluation. Instead of merging features, ICL employs multiple independent models as data quality assessors, which collectively identify reliable training samples through prediction consensus. This consensus-driven subset is used to iteratively refine the training process iteratively, enhancing generalization. Extensive experiments on multiple datasets demonstrate the strong generalization capability of our method.

Federated learning is increasingly adopted for traffic flow prediction (TFP) to enable privacy preserving collaboration across distributed sensors. However, real-world deployments are highly heterogeneous in computational capability, causing stragglers that dominate per-round latency and severely slow down model updates. Most existing approaches mitigate stragglers by suppressing or discarding slow clients, which reduce data representativeness and introduce training bias. It is a harmful trade-off for TFP where broad spatial coverage is crucial for accuracy. We propose a communication-efficient logical clustering federated learning framework (LCFed) that mitigates stragglers by logically balancing effective training time while preserving full client participation. LCFed combines a coarse-grained logical dynamic clustering algorithm (LoDynClust) to balance computational resources across clusters and reduce synchronization delays, with a fine-grained intra-cluster adaptive collaborative training mechanism (ICACT) to regulate aggregation intervals and mitigate training bias. We further provide a convergence analysis. Extensive experiments on three real-world traffic datasets show that LCFed significantly reduces training latency caused by stragglers while maintaining competitive prediction accuracy compared with state-of-the-art baselines.

Existing general-purpose forgery detection techniques fall short in military scenarios because they lack military-specific priors about how real assets are designed, manufactured, and deployed. Authentic military platforms obey strict engineering and design standards, resulting in highly regular structural layouts and characteristic material textures, whereas AI-generated forgeries often exhibit subtle violations of these constraints. To address this critical gap, we introduce SentinelFakeNet (SFNet), a novel framework specifically designed for detecting AI-generated military images. SFNet features the Military Hierarchical Perception (MHP) Module, which extracts military-relevant hierarchical representations via Cross-Level Feature Fusion (CLFF) — a mechanism that intricately combines features from varying depths of the backbone. Furthermore, to ensure robustness and adaptability to diverse generative models, we propose the Military Adaptive Test-Time Training (MATTT) strategy, which incorporates Local Consistency Verification (LCV) and Multi-Scale Signature Analysis (MSSA) as specially designed tasks. To facilitate research in this domain, we also introduce MilForgery, the first large-scale military image forensic dataset comprising 800,000 authentic and synthetically generated military-related images. Extensive experiments demonstrate that our method achieves 95.80% average accuracy, representing state-of-the-art performance. Moreover, it exhibits superior generalization capabilities on public AIGC detection benchmarks, outperforming the leading baselines by +8.47% and +6.49% on GenImage and ForenSynths in average accuracy, respectively. Our code will be available on the author’s homepage.

The generalization capability of face forgery detection systems remains a critical challenge, as their performance often deteriorates when confronted with unseen forgery techniques. To mitigate this limitation, we propose pre-trained knowledge-guided data purification for generalizable face forgery detection (PrePurify). Departing from conventional model-centric methods, this novel data-centric approach rethinks generalization through principled data purification and targeted adaptation. Face forgery data inherently contains both universal manipulation patterns and forgery-specific artifacts in varying proportions. Our key insight is that prioritizing samples with stronger universal patterns during training naturally guides the model to learn more generalizable features while reducing sensitivity to forgery-specific biases. PrePurify leverages this insight by employing a two-stage training paradigm: first, utilizing pre-trained knowledge from ImageNet to identify samples richest with universal features, then performing focused adaptation on this optimized subset. Extensive evaluations confirm PrePurify’s superior generalization across diverse face forgery domains, including emerging diffusion-based synthetic faces.

With the rapid development of Generative AI making image forgery increasingly realistic, Image Forgery Detection and Localization (IFDL) has become critical for precisely detecting and localizing forged regions. However, current IFDL methods struggle to achieve satisfactory localization precision and robustness, primarily due to their reliance on either physical traces or semantic context alone. Physical-based methods can be confused by complex textures, while semantic-based methods may overlook subtle manipulation traces. To address this, we propose PSCA-Net, a unified framework comprising a Texture Stream and a Semantic Stream that integrates physical traces and semantic contexts to capture complementary cues. Crucially, we introduce a Semantic-Texture Consistency (STC) module that explicitly calculates cross-modal conflicts, utilizing these discrepancies to verify and refine feature representations for precise localization. Extensive experiments on multiple benchmarks demonstrate that PSCA-Net not only achieves SOTA performance in both detection and localization tasks, but also exhibits superior robustness compared to existing methods.

Video Double Compression Detection (DCD) has been extensively studied under conventional video coding standards. However, the introduction of advanced coding tools in Versatile Video Coding (VVC), especially the hybrid intra prediction structure involving data-driven predictors, fundamentally alters the statistical characteristics exploited by existing DCD methods. As a result, VVC DCD under different coding parameters becomes more challenging and remains insufficiently explored. To address this problem, this paper proposes a VVC DCD method based on Hybrid Intra Prediction Modes Variation (HIPMV). The effects of VVC recompression on both conventional intra predictors and newly introduced tools, including Matrix-based Intra Prediction (MIP) and Wide Angle Intra Prediction (WAIP), are first statistically analyzed. Based on these analyses, the HIPMV feature is constructed by jointly modeling hybrid intra prediction mode variation and fine grained decision fluctuation within the MIP module. Extensive experimental results demonstrate that the proposed low dimensional HIPMV feature effectively characterizes VVC recompression traces under different coding parameter settings.

Coverless image steganography (CIS) can completely resist steganalysis algorithms as it does not modify the carriers. However, most existing CIS methods overlook the diversity of generated hash sequences, which directly impacts the practical maximum capacity and difficulty of constructing the index structure. Moreover, CIS methods typically consider robustness only under known parameters. In security, some methods require the transmission of auxiliary information along with stego images, while others reuse fixed images in the dataset, which raises potential suspicion from attackers. To address the above issues, a robust CIS method towards resisting black-box attacks is proposed. The normalized average of keypoint descriptor features is calculated by Scale-Invariant Feature Transform (SIFT) to construct the hierarchical coverless image dataset (CID). Subsequently, the mapping table between hash sequences and images in CID is established to accomplish secret information hiding and extraction. Experimental results demonstrate that the proposed CIS method achieves a higher capacity and stronger robustness than the state-of-the-art CIS methods. Notably, it is the first time to consider robustness against black-box and mixed attacks in CIS methods, making it more applicable to real-world scenarios.

The rapid proliferation of AI-Generated Content (AIGC), spanning text, images, video, and audio, has created a dual-edged sword of unprecedented creativity and significant societal risks, including misinformation and disinformation. This survey provides a comprehensive and structured overview of the current landscape of AIGC detection technologies. We begin by chronicling the evolution of generative models, from foundational GANs to state-of-the-art diffusion and transformer-based architectures. We then systematically review detection methodologies across all modalities, organizing them into a novel taxonomy of External Detection and Internal Detection. For each modality, we trace the technical progression from early feature-based methods to advanced deep learning, while also covering critical tasks like model attribution and tampered region localization. Furthermore, we survey the ecosystem of publicly available detection tools and practical applications. Finally, we distill the primary challenges facing the field–including generalization, robustness, interpretability, and the lack of universal benchmarks–and conclude by outlining key future directions, such as the development of holistic AI Safety Agents, dynamic evaluation standards, and AI-driven governance frameworks. This survey aims to provide researchers and practitioners with a clear, in-depth understanding of the state-of-the-art and critical frontiers in the ongoing endeavor to ensure a safe and trustworthy AIGC ecosystem.

The rapid advancement of facial manipulation technologies demands detection systems that can generalize to novel forgery techniques. We identify that Standard Learning, reliant on static datasets and uniform sampling, detrimentally biases models towards specific patterns tied to individual generation techniques, hindering their ability to learn general features. To overcome this, we introduce Adaptive Learning (AL) for face forgery detection, a cyclical framework that simultaneously refines both the detector model and the training data through dynamic sample selection and model optimization. AL’s efficacy hinges on identifying samples rich in generalizable forgery clues. Thus, we propose Augmentation Robustness Validation (ARV) as AL’s core purification engine. ARV exploits the stability of predictions across diverse semantic-preserving augmentations as a reliable proxy for general feature presence: samples that exhibit invariant predictions inherently contain robust manipulation traces. Integrating ARV with AL yields Adaptive Learning with Augmentation Robustness Validation (ALarv). ALarv strategically prioritizes stability-verified samples during iterative training cycles, progressively enhancing the model’s focus on transferable forensic features. Inspired by the architectural advantages of ConvNeXt, we incorporate it into ALarv, forming an effective method, ALarv-ConvNeXt. Extensive experiments demonstrate ALarv-ConvNeXt’s superior generalization performance, including emerging diffusion-based synthetic faces.

Street view images are crucial in mobile computing, autonomous driving and social media, etc.. However, the disclosure of building location information raises significant privacy concerns. Current privacy preserving solutions often suffer from poor image quality or lack control over privacy levels. To address these issues, we propose a Building Privacy Preserving Method (BPPM). We introduce a building semantic enhancement strategy to improve the quality and utility of generated images. The key idea is to leverage the location distribution and semantic information of buildings to enhance the discriminator's identification ability, thereby improving the generator's performance. Additionally, BPPM allows users to adjust both the edges and styles of generated building images to meet personalized privacy degree requirements. Extensive experiments on real-world datasets demonstrate BPPM's effectiveness against location inference attacks. Compared to SOTA approaches, BPPM outperforms with higher U-SSIM and U-IOU scores and lower P-PSNR and P-SSIM scores. The code is available at https://github.com/whwh456/ BPPM

The proliferation of hyper-realistic AI-generated images poses significant threats to digital information integrity and forensic accountability. Existing detection methodologies, however, face three critical bottlenecks: vulnerability to real-world distortions such as social media compression, inadequate specialization for socially harmful “ex-regulatory” content, and an inability to perform model attribution essential for effective governance. To address these challenges, we propose the Multi-level Feature Fusion Detection and Attribution Framework (MF2DA), a unified end-to-end pipeline designed for both high-precision detection and reliable model attribution. The core architecture synergizes an Edge Pyramid Fusion ResNet (EPF-ResNet), which captures subtle pixel-level edge artifacts, with a frozen CLIP-ViT to ensure robust semantic generalization. Furthermore, the framework is augmented by the MLLM-Guided Quality Refinement Module (MQRM), which adaptively leverages semantic-agnostic quality features to decouple generative traces from aggressive compression noise. Finally, the Dual-stream Differential Patch Attribution Network (D2PAN) extracts resilient model fingerprints by disentangling micro-textural patterns from semantic interference, thereby achieving precise generator identification. Extensive evaluations on multiple benchmarks demonstrate that MF2DA achieves state-of-the-art performance in detecting both “friendly” and “ex-regulatory” images while maintaining exceptional cross-dataset generalization. By integrating robust detection with precise attribution, this work establishes a practical and accountable forensic solution for the rapidly evolving generative AI landscape.

AI-generated video detection remains challenging under realistic compression and transcoding, where pixel-domain forensic cues are often degraded. We propose a compresseddomain spatio-temporal framework for detecting AI-generated videos. Our method is based on the observation that AI videos often produce abnormal coding behaviors during compression, including irregular coding unit partition, prediction patterns, transform statistics, and local temporal inconsistency. To exploit these traces, we construct a five-view compressed-domain representation and combine multi-branch residual feature learning with bidirectional temporal modeling. The proposed framework captures both spatial coding irregularities and short-range temporal anomalies, making it suitable for compressed video forensics in practical AIGC scenarios.

As Deepfake technology has advanced rapidly, the sharing of facial images on social media platforms faces more privacy leakage risks. Attackers can effortlessly utilize publicly available images to perform highly realistic attribute editing and face swapping at a low cost. To alleviate this threat, active defense strategies introduce imperceptible adversarial perturbations into images before they are published, thus interfering with the forgery generation process. However, existing methods suffer from poor cross-model generalization, a lack of crosstask inheritability, and extreme vulnerability to real-world image compression algorithms. To address these limitations, we propose a generalized perturbation training framework based on blackbox watermarking. Specifically, we design a Random Sampling Module and a series superposition strategy to achieve crosstask inheritance, enabling efficient adaptation to novel forgery tasks without requiring the retraining of the entire model. Furthermore, we integrate a Simulated Compression Module into the adversarial training pipeline to significantly enhance the anticompression robustness of the generated perturbations. Extensive experiments on the OSN(Online Social Network)-transmission dataset demonstrate that our framework achieves outstanding defense success rates against various face-swapping and attributeediting models, while maintaining high visual fidelity and robust protection even after undergoing complex social media compression.

The rapid development of AI-generated content technologies has enabled the creation of highly realistic synthetic videos, posing significant challenges for video authenticity verification. Most existing detection methods rely on visual artifacts in the pixel domain or temporal inconsistencies across frames, which may become less reliable as generative models continue to improve visual realism. In this paper, we propose an encoding-domain framework for AI-generated video detection by analyzing statistical characteristics derived from the video compression process. Specifically, input videos are re-encoded under multiple quantization parameter (QP) settings to extract macroblock-type distribution statistics across different compression conditions. These multi-QP statistics are organized into encoding feature matrices and modeled using a Transformer encoder with a self-supervised masked reconstruction strategy to learn robust representations, which are then fine-tuned for binary classification. Experimental results on the challenging GenVidBench benchmark demonstrate that the proposed method achieves strong detection performance and generalization under cross-source evaluation settings, highlighting the effectiveness of encoding-domain statistical analysis for AI-generated video forensics.

The goal of image forgery localization is to pinpoint altered image regions with precise pixel-level accuracy, which remains challenging due to the subtlety of forgery traces and the imbalance between tampered and background regions. With the continuous advancement of text-to-image models, this task has become more challenging. Currently, most methods address this issue by deepening and strengthening the network, resulting in a large number of parameters and a slow computation speed. This paper proposes LDB-net, a lightweight dual-branch FCN-8s framework integrated with sparse self-attention and multiscale feature enhancement for robust image forgery localization. Specifically, we design a deep RGB stream based on VGG-16 to capture high-level semantic inconsistencies and a lighter wavelet transform based frequency stream to extract low-level texture and noise clues from the frequency domain. To address the high dimensionality of the frequency stream, we introduce the Spectral-Spatial Squeeze (SSS) block which can select channels with higher information density under supervision. We replaced the native bottleneck layer of FCN with multi-scale dilated convolution to capture richer features. Additionally, we utilize sparse self-attention to extract the attention map of the frequency branch, and use this map to weight the RGB branch. We evaluated the localization performance of this model on four different datasets and found that it achieved satisfactory results with only 16.41 M trainable parameters and 105.7G FLOPs.

Deepfake has recently garnered considerable attention due to its potential threat. Recent detectors often struggle to generalize due to sensitivity to dataset-specific biases. We identify a key factor: their performance varies significantly with different normalization parameters, indicating reliance on preprocessing artifacts rather than authentic manipulation traces. To address this, we propose Normalization-Consistent Data Curation (NormCura), which selects training samples based on their prediction stability across normalization variations. NormCura first evaluates sample consistency under multiple normalization conditions, then trains only on stable samples. This filters out normalization-sensitive artifacts while retaining robust forensic patterns. Extensive cross-dataset evaluations on nine deepfake datasets demonstrate that this approach significantly improves generalization performance, including emerging diffusion-based synthetic faces, confirming that normalization consistency is an effective proxy for learning generalizable deepfake detection features.

A prevalent video manipulation technique involves upsampling the bitrate parameter without modifying the underlying video content, often implemented under the pretext of improving viewer appeal and monetization potential. Such deceptive practices, which substitute fake bitrate specifications for genuine ones, mislead audiences and platforms while directly infringing copyright protections, constituting a form of manipulation that forensic experts classify as fake bitrate video manipulation. Addressing the issue of detecting fake bitrate Versatile Video Coding (VVC) videos, an algorithm based on hierarchical feature mapping in the coding domain is proposed. We first analyze the Coding Unit (CU) partitioning and the deblocking filtering of fake bitrate VVC videos during multiple encoding processes. Then, CU partitioning and deblocking decision-mode information in the coding domain are extracted during the decoding process. By combining the positional information, the CU Difference feature map (FCD) and the Deblocking Filtering Difference feature map (FDD) are obtained through a hierarchical mapping of the encoding feature, which are further enhanced by calculating the eccentric covariance matrix. After that, the enhanced feature maps are fed into a Dual-branch Difference Perception (DDP) module to obtain frame-level detection results. By comparing with existing algorithms, the experimental results demonstrate that the proposed algorithm achieves superior detection accuracy in different scenarios, validating its effectiveness and robustness.

The growing adoption of 3D point cloud in applications like autonomous driving has heightened concerns about their vulnerability to adversarial attacks. Existing defense methods face two fundamental challenges: ineffective detection of imperceptible adversarial examples and poor restoration of severely distorted point cloud. In this paper, we present ADDR, an end-to-end defense framework that integrates Binary Geometric Feature Anomaly Detection (BGFAD) and Distorted point cloud Restoration (DPCR). BGFAD employs a dual threshold mechanism combining global distance statistics and local curvature analysis to detect both substantial and imperceptible adversarial perturbations. DPCR leverages attention enhanced feature encoding to reconstruct missing geometric structures while preserving semantic integrity through bidirectional Chamfer loss optimization. Our framework uniquely bridges traditional geometric priors with deep learning mechanisms, achieving attack-agnostic defense without classifier retraining. Extensive experiments on ModelNet40, ShapeNet and ScanObjectNN datasets demonstrate state-of-the-art performance, with about 12% higher robustness against structural attacks and $6\times $ better restoration fidelity than existing methods. ADDR maintains real-time processing capabilities while reducing adversarial success rates to <5% across diverse attacks. The code is available at https://github.com/whwh456/ADDR