Yajie Wang

He Yang, Dongyi Lv, Song Ma et al.

Dataset Condensation (DC) is a data-efficient learning paradigm that synthesizes small yet informative datasets, enabling models to match the performance of full-data training. However, recent work exposes a critical vulnerability of DC to backdoor attacks, where malicious patterns (\textit{e.g.}, triggers) are implanted into the condensation dataset, inducing targeted misclassification on specific inputs. Existing attacks always prioritize attack effectiveness and model utility, overlooking the crucial dimension of stealthiness. To bridge this gap, we propose InkDrop, which enhances the imperceptibility of malicious manipulation without degrading attack effectiveness and model utility. InkDrop leverages the inherent uncertainty near model decision boundaries, where minor input perturbations can induce semantic shifts, to construct a stealthy and effective backdoor attack. Specifically, InkDrop first selects candidate samples near the target decision boundary that exhibit latent semantic affinity to the target class. It then learns instance-dependent perturbations constrained by perceptual and spatial consistency, embedding targeted malicious behavior into the condensed dataset. Extensive experiments across diverse datasets validate the overall effectiveness of InkDrop, demonstrating its ability to integrate adversarial intent into condensed datasets while preserving model utility and minimizing detectability. Our code is available at https://github.com/lvdongyi/InkDrop.

Nan Luo, Yuanzhang Li, Yajie Wang et al.

Backdoor attacks threaten Deep Neural Networks (DNNs). Towards stealthiness, researchers propose clean-label backdoor attacks, which require the adversaries not to alter the labels of the poisoned training datasets. Clean-label settings make the attack more stealthy due to the correct image-label pairs, but some problems still exist: first, traditional methods for poisoning training data are ineffective; second, traditional triggers are not stealthy which are still perceptible. To solve these problems, we propose a two-phase and image-specific triggers generation method to enhance clean-label backdoor attacks. Our methods are (1) powerful: our triggers can both promote the two phases (i.e., the backdoor implantation and activation phase) in backdoor attacks simultaneously; (2) stealthy: our triggers are generated from each image. They are image-specific instead of fixed triggers. Extensive experiments demonstrate that our approach can achieve a fantastic attack success rate~(98.98%) with low poisoning rate~(5%), high stealthiness under many evaluation metrics and is resistant to backdoor defense methods.

Shuhao Li, Yajie Wang, Yuanzhang Li et al.

Machine Learning (ML) has made unprecedented progress in the past several decades. However, due to the memorability of the training data, ML is susceptible to various attacks, especially Membership Inference Attacks (MIAs), the objective of which is to infer the model's training data. So far, most of the membership inference attacks against ML classifiers leverage the shadow model with the same structure as the target model. However, empirical results show that these attacks can be easily mitigated if the shadow model is not clear about the network structure of the target model. In this paper, We present attacks based on black-box access to the target model. We name our attack \textbf{l-Leaks}. The l-Leaks follows the intuition that if an established shadow model is similar enough to the target model, then the adversary can leverage the shadow model's information to predict a target sample's membership.The logits of the trained target model contain valuable sample knowledge. We build the shadow model by learning the logits of the target model and making the shadow model more similar to the target model. Then shadow model will have sufficient confidence in the member samples of the target model. We also discuss the effect of the shadow model's different network structures to attack results. Experiments over different networks and datasets demonstrate that both of our attacks achieve strong performance.

Huipeng Zhou, Yu-an Tan, Yajie Wang et al.

Vision transformers (ViTs) have demonstrated impressive performance in various computer vision tasks. However, the adversarial examples generated by ViTs are challenging to transfer to other networks with different structures. Recent attack methods do not consider the specificity of ViTs architecture and self-attention mechanism, which leads to poor transferability of the generated adversarial samples by ViTs. We attack the unique self-attention mechanism in ViTs by restructuring the embedded patches of the input. The restructured embedded patches enable the self-attention mechanism to obtain more diverse patches connections and help ViTs keep regions of interest on the object. Therefore, we propose an attack method against the unique self-attention mechanism in ViTs, called Self-Attention Patches Restructure (SAPR). Our method is simple to implement yet efficient and applicable to any self-attention based network and gradient transferability-based attack methods. We evaluate attack transferability on black-box models with different structures. The result show that our method generates adversarial examples on white-box ViTs with higher transferability and higher image quality. Our research advances the development of black-box transfer attacks on ViTs and demonstrates the feasibility of using white-box ViTs to attack other black-box models.

Yinpeng Dong, Peng Chen, Senyou Deng et al.

The security of artificial intelligence (AI) is an important research area towards safe, reliable, and trustworthy AI systems. To accelerate the research on AI security, the Artificial Intelligence Security Competition (AISC) was organized by the Zhongguancun Laboratory, China Industrial Control Systems Cyber Emergency Response Team, Institute for Artificial Intelligence, Tsinghua University, and RealAI as part of the Zhongguancun International Frontier Technology Innovation Competition (https://www.zgc-aisc.com/en). The competition consists of three tracks, including Deepfake Security Competition, Autonomous Driving Security Competition, and Face Recognition Security Competition. This report will introduce the competition rules of these three tracks and the solutions of top-ranking teams in each track.

Haoran Lyu, Yajie Wang, Yu-an Tan et al.

The security of models based on new architectures such as MLP-Mixer and ViTs needs to be studied urgently. However, most of the current researches are mainly aimed at the adversarial attack against ViTs, and there is still relatively little adversarial work on MLP-mixer. We propose an adversarial attack method against MLP-Mixer called Maxwell's demon Attack (MA). MA breaks the channel-mixing and token-mixing mechanism of MLP-Mixer by controlling the part input of MLP-Mixer's each Mixer layer, and disturbs MLP-Mixer to obtain the main information of images. Our method can mask the part input of the Mixer layer, avoid overfitting of the adversarial examples to the source model, and improve the transferability of cross-architecture. Extensive experimental evaluation demonstrates the effectiveness and superior performance of the proposed MA. Our method can be easily combined with existing methods and can improve the transferability by up to 38.0% on MLP-based ResMLP. Adversarial examples produced by our method on MLP-Mixer are able to exceed the transferability of adversarial examples produced using DenseNet against CNNs. To the best of our knowledge, we are the first work to study adversarial transferability of MLP-Mixer.

Ruinan Ma, Yu-an Tan, Shangbo Wu et al.

Deep learning techniques have implemented many unconditional image generation (UIG) models, such as GAN, Diffusion model, etc. The extremely realistic images (also known as AI-Generated Content, AIGC for short) produced by these models bring urgent needs for intellectual property protection such as data traceability and copyright certification. An attacker can steal the output images of the target model and use them as part of the training data to train a private surrogate UIG model. The implementation mechanisms of UIG models are diverse and complex, and there is no unified and effective protection and verification method at present. To address these issues, we propose a two-stage unified watermark verification mechanism with high-binding effects for such models. In the first stage, we use an encoder to invisibly write the watermark image into the output images of the original AIGC tool, and reversely extract the watermark image through the corresponding decoder. In the second stage, we design the decoder fine-tuning process, and the fine-tuned decoder can make correct judgments on whether the suspicious model steals the original AIGC tool data. Experiments demonstrate our method can complete the verification work with almost zero false positive rate under the condition of only using the model output images. Moreover, the proposed method can achieve data steal verification across different types of UIG models, which further increases the practicality of the method.

Qiang He, Wentian Qu, Jiajia Dai et al.

Accurate semantic segmentation of 3D dental models is essential for digital dentistry applications such as orthodontics and dental implants. However, due to complex tooth arrangements and similarities in shape among adjacent teeth, existing methods struggle with accurate segmentation, because they often focus on local geometry while neglecting global contextual information. To address this, we propose TCATSeg, a novel framework that combines local geometric features with global semantic context. We introduce a set of sparse yet physically meaningful superpoints to capture global semantic relationships and enhance segmentation accuracy. Additionally, we present a new dataset of 400 dental models, including pre-orthodontic samples, to evaluate the generalization of our method. Extensive experiments demonstrate that TCATSeg outperforms state-of-the-art approaches.

Bo Zou, Shaofeng Wang, Hao Liu et al.

Teeth localization, segmentation, and labeling in 2D images have great potential in modern dentistry to enhance dental diagnostics, treatment planning, and population-based studies on oral health. However, general instance segmentation frameworks are incompetent due to 1) the subtle differences between some teeth' shapes (e.g., maxillary first premolar and second premolar), 2) the teeth's position and shape variation across subjects, and 3) the presence of abnormalities in the dentition (e.g., caries and edentulism). To address these problems, we propose a ViT-based framework named TeethSEG, which consists of stacked Multi-Scale Aggregation (MSA) blocks and an Anthropic Prior Knowledge (APK) layer. Specifically, to compose the two modules, we design 1) a unique permutation-based upscaler to ensure high efficiency while establishing clear segmentation boundaries with 2) multi-head self/cross-gating layers to emphasize particular semantics meanwhile maintaining the divergence between token embeddings. Besides, we collect 3) the first open-sourced intraoral image dataset IO150K, which comprises over 150k intraoral photos, and all photos are annotated by orthodontists using a human-machine hybrid algorithm. Experiments on IO150K demonstrate that our TeethSEG outperforms the state-of-the-art segmentation models on dental image segmentation.

Haozhong Sun, Zhongsen Li, Chenlin Du et al.

Quantitative magnetic resonance imaging (qMRI) requires multi-phase acqui-sition, often relying on reduced data sampling and reconstruction algorithms to accelerate scans, which inherently poses an ill-posed inverse problem. While many studies focus on measuring uncertainty during this process, few explore how to leverage it to enhance reconstruction performance. In this paper, we in-troduce PUQ, a novel approach that pioneers the use of uncertainty infor-mation for qMRI reconstruction. PUQ employs a two-stage reconstruction and parameter fitting framework, where phase-wise uncertainty is estimated during reconstruction and utilized in the fitting stage. This design allows uncertainty to reflect the reliability of different phases and guide information integration during parameter fitting. We evaluated PUQ on in vivo T1 and T2 mapping datasets from healthy subjects. Compared to existing qMRI reconstruction methods, PUQ achieved the state-of-the-art performance in parameter map-pings, demonstrating the effectiveness of uncertainty guidance. Our code is available at https://anonymous.4open.science/r/PUQ-75B2/.

Xinquan Chen, Zhenyun Yin, Shan He et al.

As large models evolve from conversational assistants into autonomous agents, challenges increasingly arise from long-horizon decision making, tool use, and real environment interaction. Existing agenticinfrastructure remain fragmented across evaluation, data management, and agent evolution, making it difficult to discover risks systematically and improve models in a continuous closed loop. In this report, we present \textbf{Safactory}, a scalable agent factory for trustworthy autonomous intelligence. Safactory integrates three tightly coupled platforms: a \textbf{Parallel Simulation Platform} for trajectory generation, a \textbf{Trustworthy Data Platform} for trajectory storage and experience extraction, and an \textbf{Autonomous Evolution Platform} for asynchronous reinforcement learning and on-policy distillation. As far as we know, Safactory is the first framework to propose a unified evolutionary pipeline for next-generation trustworthy autonomous intelligence.

Shangbo Wu, Yu-an Tan, Yajie Wang et al.

Adversarial transferability enables black-box attacks on unknown victim deep neural networks (DNNs), rendering attacks viable in real-world scenarios. Current transferable attacks create adversarial perturbation over the entire image, resulting in excessive noise that overfit the source model. Concentrating perturbation to dominant image regions that are model-agnostic is crucial to improving adversarial efficacy. However, limiting perturbation to local regions in the spatial domain proves inadequate in augmenting transferability. To this end, we propose a transferable adversarial attack with fine-grained perturbation optimization in the frequency domain, creating centralized perturbation. We devise a systematic pipeline to dynamically constrain perturbation optimization to dominant frequency coefficients. The constraint is optimized in parallel at each iteration, ensuring the directional alignment of perturbation optimization with model prediction. Our approach allows us to centralize perturbation towards sample-specific important frequency features, which are shared by DNNs, effectively mitigating source model overfitting. Experiments demonstrate that by dynamically centralizing perturbation on dominating frequency coefficients, crafted adversarial examples exhibit stronger transferability, and allowing them to bypass various defenses.

Xuhao Ren, Haotian Liang, Yajie Wang et al.

To boost the encoder stealing attack under the perturbation-based defense that hinders the attack performance, we propose a boosting encoder stealing attack with perturbation recovery named BESA. It aims to overcome perturbation-based defenses. The core of BESA consists of two modules: perturbation detection and perturbation recovery, which can be combined with canonical encoder stealing attacks. The perturbation detection module utilizes the feature vectors obtained from the target encoder to infer the defense mechanism employed by the service provider. Once the defense mechanism is detected, the perturbation recovery module leverages the well-designed generative model to restore a clean feature vector from the perturbed one. Through extensive evaluations based on various datasets, we demonstrate that BESA significantly enhances the surrogate encoder accuracy of existing encoder stealing attacks by up to 24.63\% when facing state-of-the-art defenses and combinations of multiple defenses.

Yinghan Li, Yifei Li, Jiejing Zhang et al.

It has long been a problem to arrange and execute irregular workloads on massively parallel devices. We propose a general framework for statically batching irregular workloads into a single kernel with a runtime task mapping mechanism on GPUs. We further apply this framework to Mixture-of-Experts (MoE) model inference and implement an optimized and efficient CUDA kernel. Our MoE kernel achieves up to 91% of the peak Tensor Core throughput on NVIDIA H800 GPU and 95% on NVIDIA H20 GPU.

Shanghai AI Lab, Yicheng Bao, Guanxu Chen et al.

We introduce SafeWork-R1, a cutting-edge multimodal reasoning model that demonstrates the coevolution of capabilities and safety. It is developed by our proposed SafeLadder framework, which incorporates large-scale, progressive, safety-oriented reinforcement learning post-training, supported by a suite of multi-principled verifiers. Unlike previous alignment methods such as RLHF that simply learn human preferences, SafeLadder enables SafeWork-R1 to develop intrinsic safety reasoning and self-reflection abilities, giving rise to safety `aha' moments. Notably, SafeWork-R1 achieves an average improvement of $46.54\%$ over its base model Qwen2.5-VL-72B on safety-related benchmarks without compromising general capabilities, and delivers state-of-the-art safety performance compared to leading proprietary models such as GPT-4.1 and Claude Opus 4. To further bolster its reliability, we implement two distinct inference-time intervention methods and a deliberative search mechanism, enforcing step-level verification. Finally, we further develop SafeWork-R1-InternVL3-78B, SafeWork-R1-DeepSeek-70B, and SafeWork-R1-Qwen2.5VL-7B. All resulting models demonstrate that safety and capability can co-evolve synergistically, highlighting the generalizability of our framework in building robust, reliable, and trustworthy general-purpose AI.

Yuefeng Chen, Xiaofeng Mao, Yuan He et al.

Many works have investigated the adversarial attacks or defenses under the settings where a bounded and imperceptible perturbation can be added to the input. However in the real-world, the attacker does not need to comply with this restriction. In fact, more threats to the deep model come from unrestricted adversarial examples, that is, the attacker makes large and visible modifications on the image, which causes the model classifying mistakenly, but does not affect the normal observation in human perspective. Unrestricted adversarial attack is a popular and practical direction but has not been studied thoroughly. We organize this competition with the purpose of exploring more effective unrestricted adversarial attack algorithm, so as to accelerate the academical research on the model robustness under stronger unbounded attacks. The competition is held on the TianChi platform (\url{https://tianchi.aliyun.com/competition/entrance/531853/introduction}) as one of the series of AI Security Challengers Program.

Yajie Wang, Shangbo Wu, Wenyi Jiang et al.

Deep neural networks (DNNs) have been found to be vulnerable to adversarial examples. Adversarial examples are malicious images with visually imperceptible perturbations. While these carefully crafted perturbations restricted with tight $\Lp$ norm bounds are small, they are still easily perceivable by humans. These perturbations also have limited success rates when attacking black-box models or models with defenses like noise reduction filters. To solve these problems, we propose Demiguise Attack, crafting ``unrestricted'' perturbations with Perceptual Similarity. Specifically, we can create powerful and photorealistic adversarial examples by manipulating semantic information based on Perceptual Similarity. Adversarial examples we generate are friendly to the human visual system (HVS), although the perturbations are of large magnitudes. We extend widely-used attacks with our approach, enhancing adversarial effectiveness impressively while contributing to imperceptibility. Extensive experiments show that the proposed method not only outperforms various state-of-the-art attacks in terms of fooling rate, transferability, and robustness against defenses but can also improve attacks effectively. In addition, we also notice that our implementation can simulate illumination and contrast changes that occur in real-world scenarios, which will contribute to exposing the blind spots of DNNs.