detect synthetic time series

This study addresses the challenge of detecting time series generated by diffusion models when the generator is unknown and distribution shifts are present. Existing reconstruction-based white-box detection methods struggle in this setting due to the absence of universal generative priors for time series, a fundamental distinction from the image domain. The work presents the first systematic comparison between white-box and black-box detection paradigms for this task, demonstrating that off-the-shelf black-box classifiers significantly outperform white-box approaches. Empirical results across multiple datasets show that black-box methods achieve an average F1 score of 79.2%, representing a relative improvement of 22.1%, and attain a true positive rate of 57.2% at a 1% false positive rate, thereby validating the effectiveness and superiority of the black-box strategy.

To address security risks arising from increasingly photorealistic images generated by diffusion models, this paper proposes an interpretable detection method based on the inverse multi-step noise addition process. Methodologically, it is the first to uncover systematic frequency-domain discrepancies between natural and synthetic images during reverse denoising, modeling high-frequency Fourier spectral features and introducing a temporal noise-augmented ensemble detection framework. Furthermore, a Grad-CAM–enhanced interpretability generation and refinement module enables precise localization of forged regions. The work establishes two novel benchmarks: GenHard (for high-difficulty detection) and GenExplain (for interpretability evaluation). Experiments demonstrate state-of-the-art performance—98.91% and 95.89% detection accuracy on standard and challenging samples, respectively—surpassing prior methods by ≥2.51%, with strong cross-model generalization. Code and datasets are publicly released.

The rapid advancement of generative AI—particularly GANs, diffusion models, and VAEs—has significantly intensified the risks and societal harms associated with synthetic imagery. While existing surveys predominantly focus on deepfake detection, they lack systematic coverage of multimodal digital forensics and emerging synthetic image identification techniques. To address this gap, we propose the first taxonomy of synthetic image detection methods explicitly designed for multimodal frameworks. Our survey comprehensively analyzes over 100 representative works published between 2019 and 2024, spanning key paradigms including frequency-domain analysis, texture anomaly modeling, neural artifact identification, cross-modal alignment, self-supervised pretraining, and large-model zero-shot discrimination. We consolidate more than ten mainstream public benchmarks into a structured knowledge graph, enabling rigorous algorithmic development, standardized benchmarking, and robustness evaluation. This work provides both theoretical foundations and practical guidance for advancing trustworthy multimodal forensic research.

The rapid advancement of generative diffusion models has intensified challenges in detecting deepfakes and tracing their origins, as existing supervised detectors suffer from poor cross-generator generalization, heavy reliance on labeled data, and frequent retraining requirements. To address these limitations, we propose FRIDA—a novel framework that leverages frozen internal activation features from pre-trained diffusion models, enabling universal deepfake detection and generator attribution without fine-tuning. FRIDA exploits inherent generator-specific patterns embedded in diffusion features: it employs a k-nearest neighbors classifier for binary forgery detection and a lightweight neural network for fine-grained source identification. Extensive experiments demonstrate that FRIDA achieves state-of-the-art performance on cross-generator detection and significantly outperforms existing supervised methods in generator attribution accuracy. Moreover, it exhibits strong generalization across unseen generators, offers interpretable decisions via feature-space proximity, and maintains deployment efficiency with minimal computational overhead.

Existing AI-generated image detectors exhibit insufficient generalization to unseen text-to-image diffusion models. Method: We introduce the largest forgery image benchmark to date—comprising 4,803 distinct diffusion models and 2.7 million highly diverse images—uniquely integrating both open-source and commercial models across architectures and training configurations. We propose an end-to-end framework encompassing automated web crawling, heterogeneous model integration, and unified detector training. Contribution/Results: Empirical analysis reveals that both model count and architectural diversity confer dual positive gains in detector generalization. Our approach achieves significantly higher detection accuracy on previously unseen generators than current state-of-the-art methods, empirically validating data diversity as an effective pathway to enhanced generalization in AI-generated image detection.

This work addresses the limited generalization of existing synthetic image detection methods, which typically rely on large amounts of labeled data. The authors propose the first fully training-free detection framework: it leverages the pretrained Noiseprint++ model to extract noise residuals and combines them with multi-scale features from a frozen Vision Transformer, adaptively fused for robust representation. Only a small set of authentic images is required to initialize cluster centroids, enabling unsupervised K-Means clustering to effectively distinguish real from synthetic images. Evaluated on four benchmark datasets, the method achieves an average accuracy of 82.2%, significantly outperforming current state-of-the-art approaches and demonstrating exceptional generalization—particularly on images generated by previously unseen diffusion models.

This work addresses the challenge that existing unsupervised methods struggle to reliably detect subtle and noisy anomalies in complex time series, often being misled by noise in normal samples and missing near-normal anomalies. To overcome this limitation, we propose a novel unsupervised anomaly detection framework that integrates active learning: it enhances temporal dependency modeling through a masked time series reconstruction feedback mechanism and employs a minimax optimization strategy to differentially treat normal and anomalous samples, thereby improving robustness against noise and weak anomalies. Extensive experiments across four multivariate time series datasets and seven backbone models demonstrate that our method achieves an average AUC improvement of 12.39%, significantly outperforming current unsupervised approaches.

Existing research on detecting reward hacking in code generation relies heavily on synthetic data, yet its effectiveness in real-world scenarios remains unclear. This work systematically investigates the discrepancies between synthetic and real-world reward hacking behaviors and introduces a scalable method for collecting authentic hacking trajectories. By enhancing the GRPO algorithm with conflicting unit test injection and a “resample-until-hack” mechanism, the authors enable large-scale acquisition of real hacking trajectories. Leveraging this dataset, they establish a comparative analysis framework that reveals, for the first time, a significant gap between synthetic data and actual reward hacking behaviors. Their findings demonstrate that detectors trained solely on synthetic data exhibit limited generalization, whereas those trained on real trajectories show substantially stronger capability in identifying unseen types of reward hacking.

Existing methods for detecting images generated by diffusion models rely on time-consuming reconstruction and exhibit poor generalization. This work proposes FIND, a novel approach that, for the first time, trains a lightweight binary classifier by perturbing real images with Gaussian noise and labeling them as synthetic samples. FIND directly captures the intrinsic distributional discrepancy between real and synthetic images in terms of their difficulty in Gaussian fitting, without requiring image reconstruction or priors specific to any generative model. The method establishes an end-to-end efficient detection framework that achieves strong performance on the GenImage benchmark, improving detection accuracy by 11.7% while operating 126 times faster than current state-of-the-art approaches—demonstrating a significant advance in balancing universality, efficiency, and practicality.

This study addresses sequential multi-stream detection under the constraint that only one data stream can be observed at each time step, with the goal of simultaneously controlling global false alarm and missed detection probabilities while minimizing detection delay. To this end, the work introduces a novel optimality criterion based on the expected order statistics of detection times and proposes an active sampling strategy—dubbed “follow-the-leader”—that integrates exploration and exploitation mechanisms. Theoretical analysis demonstrates that the proposed strategy achieves asymptotic optimality for all such criteria as error probabilities vanish. Numerical experiments further confirm its superior finite-sample performance compared to existing methods and show that it closely approaches the performance of an ideal oracle policy that has full knowledge of the anomalous streams.