Text-to-image generative models have achieved impressive fidelity and diversity, but can inadvertently produce unsafe or undesirable content due to implicit biases embedded in large-scale training datasets. Existing concept erasure methods, whether text-only or image-assisted, face trade-offs: textual approaches often fail to fully suppress concepts, while naive image-guided methods risk over-erasing unrelated content. We propose TICoE, a text-image Collaborative Erasing framework that achieves precise and faithful concept removal through a continuous convex concept manifold and hierarchical visual representation learning. TICoE precisely removes target concepts while preserving unrelated semantic and visual content. To objectively assess the quality of erasure, we further introduce a fidelity-oriented evaluation strategy that measures post-erasure usability. Experiments on multiple benchmarks show that TICoE surpasses prior methods in concept removal precision and content fidelity, enabling safer, more controllable text-to-image generation. Our code is available at https://github.com/OpenAscent-L/TICoE.git

Text-to-image (T2I) diffusion models have achieved remarkable success in image synthesis, but their reliance on large-scale data and open ecosystems introduces serious backdoor security risks. Existing defenses, particularly input-level methods, are more practical for deployment but often rely on observable anomalies that become unreliable under stealthy, semantics-preserving trigger designs. As modern backdoor attacks increasingly embed triggers into natural inputs, these methods degrade substantially, raising a critical question: can more stable, implicit, and trigger-agnostic differences between benign and backdoor inputs be exploited for detection? In this work, we address this challenge from an active probing perspective. We introduce controlled scaling perturbations on cross-attention and uncover a novel phenomenon termed Cross-Attention Scaling Response Divergence (CSRD), where benign and backdoor inputs exhibit systematically different response evolution patterns across denoising steps. Building on this insight, we propose SET, an input-level backdoor detection framework that constructs response-offset features under multi-scale perturbations and learns a compact benign response space from a small set of clean samples. Detection is then performed by measuring deviations from this learned space, without requiring prior knowledge of the attack or access to model training. Extensive experiments demonstrate that SET consistently outperforms existing baselines across diverse attack methods, trigger types, and model settings, with particularly strong gains under stealthy implicit-trigger scenarios. Overall, SET improves AUROC by 9.1% and ACC by 6.5% over the best baseline, highlighting its effectiveness and robustness for practical deployment.

Mixture-of-Experts (MoE) has demonstrated strong performance in video understanding tasks, yet its adversarial robustness remains underexplored. Existing attack methods often treat MoE as a unified architecture, overlooking the independent and collaborative weaknesses of key components such as routers and expert modules. To fill this gap, we propose Temporal Lipschitz-Guided Attacks (TLGA) to thoroughly investigate component-level vulnerabilities in video MoE models. We first design attacks on the router, revealing its independent weaknesses. Building on this, we introduce Joint Temporal Lipschitz-Guided Attacks (J-TLGA), which collaboratively perturb both routers and experts. This joint attack significantly amplifies adversarial effects and exposes the Achilles' Heel (collaborative weaknesses) of the MoE architecture. Based on these insights, we further propose Joint Temporal Lipschitz Adversarial Training (J-TLAT). J-TLAT performs joint training to further defend against collaborative weaknesses, enhancing component-wise robustness. Our framework is plug-and-play and reduces inference cost by more than 60% compared with dense models. It consistently enhances adversarial robustness across diverse datasets and architectures, effectively mitigating both the independent and collaborative weaknesses of MoE.

The generalization problem remains a critical challenge in face forgery detection. Some researches have discovered that ``a backdoor path" in the representations from forgery-irrelevant information to labels induces biased learning, thereby hindering the generalization. In this paper, these forgery-irrelevant information are collectively termed spurious correlations factors. Previous methods predominantly focused on identifying concrete, specific spurious correlation and designing corresponding solutions to address them. However, spurious correlations arise from unobservable confounding factors, making it impractical to identify and address each one individually. To address this, we propose an intervention paradigm for representation space. Instead of tracking and blocking various instance-level spurious correlation one by one, we uniformly model them as a low-rank subspace and intervene in them. Specifically, we decompose spurious correlation features into a low-rank subspace via orthogonal low-rank projection, subsequently removing this subspace from the original representation and training its orthogonal complement to capture forgery-related features. This low-rank projection removal effectively eliminates spurious correlation factors, ensuring that classification decision is based on authentic forgery cues. With only 0.43M trainable parameters, our method achieves state-of-the-art performance across several benchmarks, demonstrating excellent robustness and generalization.

Deepfake detection still faces significant challenges in cross-dataset and real-world complex scenarios. The root cause lies in the high diversity of artifact distributions introduced by different forgery methods, while pretrained models tend to disrupt their original general semantic structures when adapting to new artifacts. Existing approaches usually rely on indiscriminate global parameter updates or introduce additional supervision signals, making it difficult to effectively model diverse forgery artifacts while preserving semantic stability. To address these issues, this paper proposes a deepfake detection method based on Multi-Artifact Subspaces and selective layer masks (MASM), which explicitly decouples semantic representations from artifact representations and constrains the fitting strength of artifact subspaces, thereby improving generalization robustness in cross-dataset scenarios. Specifically, MASM applies singular value decomposition to model weights, partitioning pretrained weights into a stable semantic principal subspace and multiple learnable artifact subspaces. This design enables decoupled modeling of different forgery artifact patterns while preserving the general semantic subspace. On this basis, a selective layer mask strategy is introduced to adaptively regulate the update behavior of corresponding network layers according to the learning state of each artifact subspace, suppressing overfitting to any single forgery characteristic. Furthermore, orthogonality constraints and spectral consistency constraints are imposed to jointly regularize multiple artifact subspaces, guiding them to learn complementary and diverse artifact representations while maintaining a stable overall spectral structure.

With the rapid advancement of generative models, visually realistic AI-generated images have become increasingly difficult to distinguish from authentic ones, posing severe threats to social trust and information integrity. Consequently, there is an urgent need for efficient and truly explainable image forensic methods. Recent detection paradigms have shifted towards explainable forensics. However, state-of-the-art approaches primarily rely on post-hoc rationalizations or visual discrimination, lacking a verifiable chain of evidence. This reliance on surface-level pattern matching limits the generation of causally grounded explanations and often results in poor generalization. To bridge this critical gap, we introduce \textbf{REVEAL-Bench}, the first reasoning-enhanced multimodal benchmark for AI-generated image detection that is explicitly structured around a chain-of-evidence derived from multiple lightweight expert models, then records step-by-step reasoning traces and evidential justifications. Building upon this dataset, we propose \textbf{REVEAL} (\underline{R}easoning-\underline{e}nhanced Forensic E\underline{v}id\underline{e}nce \underline{A}na\underline{l}ysis), an effective and explainable forensic framework that integrates detection with a novel expert-grounded reinforcement learning. Our reward mechanism is specially tailored to jointly optimize detection accuracy, explanation fidelity, and logical coherence grounded in explicit forensic evidence, enabling REVEAL to produce fine-grained, interpretable, and verifiable reasoning chains alongside its detection outcomes. Extensive experimental results demonstrate that REVEAL significantly enhances detection accuracy, explanation fidelity, and robust cross-model generalization, benchmarking a new state of the art for explainable image forensics.

The rapid advancement of generative models has made real and synthetic images increasingly indistinguishable. Although extensive efforts have been devoted to detecting AI-generated images, out-of-distribution generalization remains a persistent challenge. We trace this weakness to spurious shortcuts exploited during training and we also observe that small feature-space perturbations can mitigate shortcut dominance. To address this problem in a more controllable manner, we propose the Positive-Incentive Noise for CLIP (PiN-CLIP), which jointly trains a noise generator and a detection network under a variational positive-incentive principle. Specifically, we construct positive-incentive noise in the feature space via cross-attention fusion of visual and categorical semantic features. During optimization, the noise is injected into the feature space to fine-tune the visual encoder, suppressing shortcut-sensitive directions while amplifying stable forensic cues, thereby enabling the extraction of more robust and generalized artifact representations. Comparative experiments are conducted on an open-world dataset comprising synthetic images generated by 42 distinct generative models. Our method achieves new state-of-the-art performance, with notable improvements of 5.4 in average accuracy over existing approaches.

The rapid progress of generative models such as GANs and diffusion models has led to the widespread proliferation of AI-generated images, raising concerns about misinformation, privacy violations, and trust erosion in digital media. Although large-scale multimodal models like CLIP offer strong transferable representations for detecting synthetic content, fine-tuning them often induces catastrophic forgetting, which degrades pre-trained priors and limits cross-domain generalization. To address this issue, we propose the Distillation-guided Gradient Surgery Network (DGS-Net), a novel framework that preserves transferable pre-trained priors while suppressing task-irrelevant components. Specifically, we introduce a gradient-space decomposition that separates harmful and beneficial descent directions during optimization. By projecting task gradients onto the orthogonal complement of harmful directions and aligning with beneficial ones distilled from a frozen CLIP encoder, DGS-Net achieves unified optimization of prior preservation and irrelevant suppression. Extensive experiments on 50 generative models demonstrate that our method outperforms state-of-the-art approaches by an average margin of 6.6, achieving superior detection performance and generalization across diverse generation techniques.

The rapid progress of generative models, such as GANs and diffusion models, has facilitated the creation of highly realistic images, raising growing concerns over their misuse in security-sensitive domains. While existing detectors perform well under known generative settings, they often fail to generalize to unknown generative models, especially when semantic content between real and fake images is closely aligned. In this paper, we revisit the use of CLIP features for AI-generated image detection and uncover a critical limitation: the high-level semantic information embedded in CLIP's visual features hinders effective discrimination. To address this, we propose NS-Net, a novel detection framework that leverages NULL-Space projection to decouple semantic information from CLIP's visual features, followed by contrastive learning to capture intrinsic distributional differences between real and generated images. Furthermore, we design a Patch Selection strategy to preserve fine-grained artifacts by mitigating semantic bias caused by global image structures. Extensive experiments on an open-world benchmark comprising images generated by 40 diverse generative models show that NS-Net outperforms existing state-of-the-art methods, achieving a 7.4\% improvement in detection accuracy, thereby demonstrating strong generalization across both GAN- and diffusion-based image generation techniques.