The massive scale of data and computation required for training Multimodal Large Language Models (MLLMs) has fueled the rise of Fine-Tuning as a Service (FTaaS), enabling users to rapidly customize models for diverse real-world tasks. While FTaaS democratizes access to advanced multimodal intelligence, it also introduces serious security concerns, particularly backdoor attacks. In this work, we systematically analyze backdoor vulnerabilities in MLLMs under the FTaaS paradigm, revealing two key phenomena: (1) markedly reduced sensitivity to textual variations when a visual trigger is present, and (2) abnormally stable model confidence even under strong semantic perturbations. Building on these insights, we propose Trap on Text (ToT), a novel inference-time backdoor detection framework. ToT applies controlled semantic perturbations to textual prompts and jointly analyzes the semantic consistency and confidence drift of the model’s responses, enabling robust detection of backdoor activations without requiring model parameters, architectures or clean reference data. Extensive experiments across architectures and datasets show that ToT achieves strong attack mitigation and preserves clean accuracy, offering a practical solution for safeguarding FTaaS workflows.

Despite the integration of safety alignment and external filters, text-to-image (T2I) generative systems are still susceptible to producing harmful content, such as sexual or violent imagery. This raises serious concerns about unintended exposure and potential misuse. Red teaming, which aims to proactively identify diverse prompts that can elicit unsafe outputs from the T2I system, is increasingly recognized as an essential method for assessing and improving safety before real-world deployment. However, existing automated red teaming approaches often treat prompt discovery as an isolated, prompt-level optimization task, which limits their scalability, diversity, and overall effectiveness. To bridge this gap, in this paper, we propose DREAM, a scalable red teaming framework to automatically uncover diverse problematic prompts from a given T2I system. Unlike prior work that optimizes prompts individually, DREAM directly models the probabilistic distribution of the target system's problematic prompts, which enables explicit optimization over both effectiveness and diversity, and allows efficient large-scale sampling after training. To achieve this without direct access to representative training samples, we draw inspiration from energy-based models and reformulate the objective into a simple and tractable form. We further introduce GC-SPSA, an efficient optimization algorithm that provides stable gradient estimates through the long and potentially non-differentiable T2I pipeline. During inference, we also propose a diversity-aware sampling strategy to enhance prompt variety. The effectiveness of DREAM is validated through extensive experiments, demonstrating state-of-the-art performance across a wide range of T2I models and safety filters in terms of both prompt success rate and diversity. Our code is available at https://github.com/AntigoneRandy/DREAM

Large web-scale datasets have driven the rapid advancement of pre-trained language models (PLMs), but unauthorized data usage has raised serious copyright concerns. Existing dataset ownership verification (DOV) methods typically assume that watermarks remain stable during inference; however, this assumption often fails under natural noise and adversary-crafted perturbations. We propose the first certified dataset ownership verification method for PLMs under a gray-box setting (i.e., the defender can only query the suspicious model but is aware of its input representation module), based on dual-space smoothing (i.e., DSSmoothing). To address the challenges of text discreteness and semantic sensitivity, DSSmoothing introduces continuous perturbations in the embedding space to capture semantic robustness and applies controlled token reordering in the permutation space to capture sequential robustness. DSSmoothing consists of two stages: in the first stage, triggers are collaboratively embedded in both spaces to generate norm-constrained and robust watermarked datasets; in the second stage, randomized smoothing is applied in both spaces during verification to compute the watermark robustness (WR) of suspicious models and statistically compare it with the principal probability (PP) values of a set of benign models. Theoretically, DSSmoothing provides provable robustness guarantees for dataset ownership verification by ensuring that WR consistently exceeds PP under bounded dual-space perturbations. Extensive experiments on multiple representative web datasets demonstrate that DSSmoothing achieves stable and reliable verification performance and exhibits robustness against potential adaptive attacks. Our code is available at https://github.com/NcepuQiaoTing/DSSmoothing.

The substantial investment required to develop Large Language Models (LLMs) makes them valuable intellectual property, raising significant concerns about copyright protection. LLM fingerprinting has emerged as a key technique to address this, which aims to verify a model's origin by extracting an intrinsic, unique signature (a ''fingerprint'') and comparing it to that of a source model to identify illicit copies. However, existing black-box fingerprinting methods often fail to generate distinctive LLM fingerprints. This ineffectiveness arises because black-box methods typically rely on model outputs, which lose critical information about the model's unique parameters due to the usage of non-linear functions. To address this, we first leverage Fisher Information Theory to formally demonstrate that the gradient of the model's input is a more informative feature for fingerprinting than the output. Based on this insight, we propose ZeroPrint, a novel method that approximates these information-rich gradients in a black-box setting using zeroth-order estimation. ZeroPrint overcomes the challenge of applying this to discrete text by simulating input perturbations via semantic-preserving word substitutions. This operation allows ZeroPrint to estimate the model's Jacobian matrix as a unique fingerprint. Experiments on the standard benchmark show ZeroPrint achieves a state-of-the-art effectiveness and robustness, significantly outperforming existing black-box methods.

Private data, when published online, may be collected by unauthorized parties to train deep neural networks (DNNs). To protect privacy, defensive noises can be added to original samples to degrade their learnability by DNNs. Recently, unlearnable examples (Huang et al., 2021) are proposed to minimize the training loss such that the model learns almost nothing. However, raw data are often pre-processed before being used for training, which may restore the private information of protected data. In this paper, we reveal the data privacy violation induced by data augmentation, a commonly used data pre-processing technique to improve model generalization capability, which is the first of its kind as far as we are concerned. We demonstrate that data augmentation can significantly raise the accuracy of the model trained on unlearnable examples from 21.3% to 66.1%. To address this issue, we propose a defense framework, dubbed Armor, to protect data privacy from potential breaches of data augmentation. To overcome the difficulty of having no access to the model training process, we design a non-local module-assisted surrogate model that better captures the effect of data augmentation. In addition, we design a surrogate augmentation selection strategy that maximizes distribution alignment between augmented and non-augmented samples, to choose the optimal augmentation strategy for each class. We also use a dynamic step size adjustment algorithm to enhance the defensive noise generation process. Extensive experiments are conducted on 4 datasets and 5 data augmentation methods to verify the performance of Armor. Comparisons with 6 state-of-the-art defense methods have demonstrated that Armor can preserve the unlearnability of protected private data under data augmentation. Armor reduces the test accuracy of the model trained on augmented protected samples by as much as 60% more than baselines. We also show that Armor is robust to adversarial training. We will open-source our codes upon publication.

Backdoor attacks intend to inject hidden backdoor into the deep neural networks (DNNs), such that the predictions of infected models will be maliciously changed if the hidden backdoor is activated by the attacker-specified trigger pattern. Since the infected models behave normally on predicting benign samples, the backdoor attack is stealthy and therefore a serious threat to practical applications of DNNs. Currently, most existing backdoor attacks adopted the setting of static trigger, i.e ., triggers across the training and testing images follow the same appearance and are located in the same area. In this paper, we revisit this attack paradigm by analyzing trigger characteristics. We demonstrate that this attack paradigm is vulnerable when the trigger in testing images is not consistent with the one used for training. As such, those attacks are far less effective in the physical world, where the location and appearance of the trigger contained in digitized test samples may be different from that of the one used for training. Besides, we introduce a plug-in attack enhancement module during training, inspired by the expectation over transformation (EOT), to alleviate such inconsistency vulnerability. Based on this plug-in module, we also reveal that the widely adopted data augmentation may exacerbate the security risks of backdoor attacks, although it can enhance model performance. Moreover, we evaluate our methods on multiple benchmark datasets to verify their effectiveness. We hope that our work could inspire more explorations on the properties of backdoor attacks, to facilitate the design of more robust and secure DNNs.

Watermarking for diffusion images has drawn considerable attention due to the widespread use of text-to-image diffusion models and the increasing need for their copyright protection. Recently, advanced watermarking techniques, such as Tree-Ring, integrate watermarks by embedding traceable patterns (e.g., Rings) into the latent distribution during the diffusion process. Such methods disrupt the original semantics of the generated images due to the inevitable distribution shift caused by the watermarks, thereby limiting their practicality, particularly in digital art creation. In this work, we present Semantic-aware Pivotal Tuning Watermarks (PT-Mark), a novel invisible watermarking method that preserves both the semantics of diffusion images and the traceability of the watermark. PT-Mark preserves the original semantics of the watermarked image by gradually aligning the generation trajectory with the original (pivotal) trajectory while maintaining the traceable watermarks during whole diffusion denoising process. To achieve this, we first compute the salient regions of the watermark at each diffusion denoising step as a spatial prior to identify areas that can be aligned without disrupting the watermark pattern. Guided by the region, we then introduce an additional pivotal tuning branch that optimizes the null-text embedding to align the semantics while preserving the watermarks. Extensive evaluations demonstrate that PT-Mark can preserve the original semantics of the diffusion images while integrating robust watermarks. It achieves a 10% improvement in the performance of semantic preservation compared to state-of-the-art watermarking methods, while also showing comparable robustness against real-world perturbations and four times greater efficiency. The code is available athttps://github.com/annpion/PT-Mark.

Text-to-image (T2I) diffusion models enable high-quality image generation conditioned on textual prompts. However, fine-tuning these pre-trained models for personalization raises concerns about unauthorized dataset usage. To address this issue, dataset ownership verification (DOV) has recently been proposed, which embeds watermarks into fine-tuning datasets via backdoor techniques. These watermarks remain dormant on benign samples but produce owner-specified outputs when triggered. Despite its promise, the robustness of DOV against copyright evasion attacks (CEA) remains unexplored. In this paper, we investigate how adversaries can circumvent these mechanisms, enabling models trained on watermarked datasets to bypass ownership verification. We begin by analyzing the limitations of potential attacks achieved by backdoor removal, including TPD and T2IShield. In practice, TPD suffers from inconsistent effectiveness due to randomness, while T2IShield fails when watermarks are embedded as local image patches. To this end, we introduce CEAT2I, the first CEA specifically targeting DOV in T2I diffusion models. CEAT2I consists of three stages: (1) motivated by the observation that T2I models converge faster on watermarked samples with respect to intermediate features rather than training loss, we reliably detect watermarked samples; (2) we iteratively ablate tokens from the prompts of detected samples and monitor feature shifts to identify trigger tokens; and (3) we apply a closed-form concept erasure method to remove the injected watermarks. Extensive experiments demonstrate that CEAT2I effectively evades state-of-the-art DOV mechanisms while preserving model performance. The code is available at https://github.com/csyufei/CEAT2I.

In recent years, RAG has emerged as a key paradigm for enhancing large language models (LLMs). By integrating externally retrieved information, RAG alleviates issues like outdated knowledge and, crucially, insufficient domain expertise. While effective, RAG introduces new risks of external data extraction attacks (EDEAs), where sensitive or copyrighted data in its knowledge base may be extracted verbatim. These risks are particularly acute when RAG is used to customize specialized LLM applications with private knowledge bases. Despite initial studies exploring these risks, they often lack a formalized framework, robust attack performance, and comprehensive evaluation, leaving critical questions about real-world EDEA feasibility unanswered. In this paper, we present the first comprehensive study to formalize EDEAs against retrieval-augmented LLMs. We first formally define EDEAs and propose a unified framework decomposing their design into three components: extraction instruction, jailbreak operator, and retrieval trigger, under which prior attacks can be considered instances within our framework. Guided by this framework, we develop SECRET: a Scalable and EffeCtive exteRnal data Extraction aTtack. Specifically, SECRET incorporates (1) an adaptive optimization process using LLMs as optimizers to generate specialized jailbreak prompts for EDEAs, and (2) cluster-focused triggering, an adaptive strategy that alternates between global exploration and local exploitation to efficiently generate effective retrieval triggers. Extensive evaluations across 4 models reveal that SECRET significantly outperforms previous attacks, and is highly effective against all 16 tested RAG instances. Notably, SECRET successfully extracts 35% of the data from RAG powered by Claude 3.7 Sonnet for the first time, whereas other attacks yield 0% extraction. Our findings call for attention to this emerging threat.

Deep neural networks (DNNs) are vulnerable to backdoor attacks, where an attacker manipulates a small portion of the training data to implant hidden backdoors into the model. The compromised model behaves normally on clean samples but misclassifies backdoored samples into the attacker-specified target class, posing a significant threat to real-world DNN applications. Currently, several empirical defense methods have been proposed to mitigate backdoor attacks, but they are often bypassed by more advanced backdoor techniques. In contrast, certified defenses based on randomized smoothing have shown promise by adding random noise to training and testing samples to counteract backdoor attacks. In this paper, we reveal that existing randomized smoothing defenses implicitly assume that all samples are equidistant from the decision boundary. However, it may not hold in practice, leading to suboptimal certification performance. To address this issue, we propose a certified backdoor defense method with sample-specific smoothing noises, termed Cert-SSBD. Cert-SSBD first employs stochastic gradient ascent to optimize the noise magnitude for each sample, ensuring a sample-specific noise level that is then applied to multiple poisoned training sets to retrain several smoothed models. After that, Cert-SSBD aggregates the predictions of multiple smoothed models to generate the final robust prediction. In particular, in this case, existing certification methods become inapplicable since the optimized noise varies across different samples. To conquer this challenge, we introduce a storage-update-based certification method, which dynamically adjusts each sample’s certification region to improve certification performance. We conduct extensive experiments on multiple benchmark datasets, demonstrating the effectiveness of our proposed method. Our code is available at https://github.com/NcepuQiaoTing/Cert-SSBD.

Model fingerprinting has emerged as a crucial mechanism for safeguarding the intellectual property of open-source models, offering a non-intrusive approach that requires no modifications to the protected model. However, our analysis reveals that existing fingerprinting techniques are fundamentally vulnerable to false claim attacks, wherein adversaries can fraudulently assert ownership over independent third-party models. We demonstrate that this vulnerability stems from the untargeted nature of current methods, which evaluate model similarity based on arbitrary sample outputs rather than alignment with a specific, predefined reference. To mitigate this vulnerability, we introduce FIT-Print, a targeted fingerprinting paradigm that actively counters false claim attacks. Specifically, FIT-Print leverages optimization to transform the fingerprint into a verifiable, targeted signature. Building upon this foundation, we propose two black-box fingerprinting methods, the bit-wise FIT-ModelDiff and the list-wise FIT-LIME, which utilize output distances and feature attributions as robust model signatures, respectively. Extensive evaluations across benchmark models and datasets show that our framework perfectly neutralizes false claim attacks (100% defense success rate) and eliminates false alarms on independent models (0.0%), all while maintaining a 100% ownership verification rate against diverse model reuse techniques.

Qingjie Zhang, Di Wang, Haoting Qian, Yiming Li, Tianwei Zhang, Minlie Huang, Ke Xu, Hewu Li, Liu Yan, Han Qiu. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Recent advances in large-scale text-to-image (T2I) diffusion models have enabled a variety of downstream applications. As T2I models require extensive resources for training, they constitute highly valued intellectual property (IP) for their legitimate owners, yet making them incentive targets for unauthorized fine-tuning by adversaries seeking to leverage these models for customized, usually profitable applications. Existing IP protection methods for diffusion models generally involve embedding watermark patterns and then verifying ownership through generated outputs examination, or inspecting the model’s feature space. However, these techniques are inherently ineffective in practical scenarios when the watermarked model undergoes fine-tuning, and the feature space is inaccessible during verification (i.e., black-box setting). The model is prone to forgetting the previously learned watermark knowledge when it adapts to a new task. To address this challenge, we propose SleeperMark, a novel framework designed to embed resilient watermarks into T2I diffusion models. SleeperMark explicitly guides the model to disentangle the watermark information from the semantic concepts it learns, allowing the model to retain the embedded watermark while continuing to be adapted to new downstream tasks. Our extensive experiments demonstrate the effectiveness of SleeperMark across various types of diffusion models, including latent diffusion models (e.g., Stable Diffusion) and pixel diffusion models (e.g., DeepFloyd-IF), showing robustness against downstream fine-tuning and various attacks at both the image and model levels, with minimal impact on the model’s generative capability. The code is available at https://github.com/taco-group/SleeperMark.

Federated learning is usually regarded as a privacy-preserving training paradigm for it enables multiple clients to participate in a training task without sharing their private data. However, recent studies revealed that a malicious server can still recover private data from the victim clients based on the shared gradients via deep leakage from gradients (DLG). Currently, almost all DLG attacks are designed based on the average loss, leading to a significant decrease in attack efficiency when the batch size is greater than 1. In this paper, we revisit DLG attacks from the perspective of the loss function. We reveal that not all samples in the target batch are equally susceptible to DLG attacks: the sample with the highest loss value tends to be easily recovered by DLG attacks. Based on these observations, we propose a simple yet effective DLG method under practical FL settings. Specifically, the adversaries can enhance the effectiveness of DLG by perturbing the global model through finetuning it with a few mislabeled samples (dubbed ‘Anti-FT’). Extensive experiments are conducted on benchmark datasets, which verify the effectiveness of our method and its resistance to potential defenses. The codes are available at https://github.com/zlh-thu/anti-finetune.

Backdoor attacks on deep neural networks (DNNs) have emerged as a significant security threat, allowing adversaries to implant hidden malicious behaviors during the model training phase. Pre-processing-based defense, which is one of the most important defense paradigms, typically focuses on input transformations or backdoor trigger inversion (BTI) to deactivate or eliminate embedded backdoor triggers during the inference process. However, these methods suffer from inherent limitations: transformation-based defenses often fail to balance model utility and defense performance, while BTI-based defenses struggle to accurately reconstruct trigger patterns without prior knowledge. In this paper, we propose REFINE, an inversion-free backdoor defense method based on model reprogramming. REFINE consists of two key components: \textbf{(1)} an input transformation module that disrupts both benign and backdoor patterns, generating new benign features; and \textbf{(2)} an output remapping module that redefines the model's output domain to guide the input transformations effectively. By further integrating supervised contrastive loss, REFINE enhances the defense capabilities while maintaining model utility. Extensive experiments on various benchmark datasets demonstrate the effectiveness of our REFINE and its resistance to potential adaptive attacks.

Artificial Intelligence Generated Content (AIGC) has advanced significantly, particularly with the development of video generation models such as text-to-video (T2V) models and image-to-video (I2V) models. However, like other AIGC types, video generation requires robust content control. A common approach is to embed watermarks, but most research has focused on images, with limited attention given to videos. Traditional methods, which embed watermarks frame-by-frame in a post-processing manner, often degrade video quality. In this paper, we propose VideoShield, a novel watermarking framework specifically designed for popular diffusion-based video generation models. Unlike post-processing methods, VideoShield embeds watermarks directly during video generation, eliminating the need for additional training. To ensure video integrity, we introduce a tamper localization feature that can detect changes both temporally (across frames) and spatially (within individual frames). Our method maps watermark bits to template bits, which are then used to generate watermarked noise during the denoising process. Using DDIM Inversion, we can reverse the video to its original watermarked noise, enabling straightforward watermark extraction. Additionally, template bits allow precise detection for potential spatial and temporal modification. Extensive experiments across various video models (both T2V and I2V models) demonstrate that our method effectively extracts watermarks and detects tamper without compromising video quality. Furthermore, we show that this approach is applicable to image generation models, enabling tamper detection in generated images as well. Codes and models are available at https://github.com/hurunyi/VideoShield.

Backdoor unalignment attacks against Large Language Models (LLMs) enable the stealthy compromise of safety alignment using a hidden trigger while evading normal safety auditing. These attacks pose significant threats to the applications of LLMs in the real-world Large Language Model as a Service (LLMaaS) setting, where the deployed model is a fully black-box system that can only interact through text. Furthermore, the sample-dependent nature of the attack target exacerbates the threat. Instead of outputting a fixed label, the backdoored LLM follows the semantics of any malicious command with the hidden trigger, significantly expanding the target space. In this paper, we introduce BEAT, a black-box defense that detects triggered samples during inference to deactivate the backdoor. It is motivated by an intriguing observation (dubbed the **probe concatenate effect**), where concatenated triggered samples significantly reduce the refusal rate of the backdoored LLM towards a malicious probe, while non-triggered samples have little effect. Specifically, BEAT identifies whether an input is triggered by measuring the degree of distortion in the output distribution of the probe before and after concatenation with the input. Our method addresses the challenges of sample-dependent targets from an opposite perspective. It captures the impact of the trigger on the refusal signal (which is sample-independent) instead of sample-specific successful attack behaviors. It overcomes black-box access limitations by using multiple sampling to approximate the output distribution. Extensive experiments are conducted on various backdoor attacks and LLMs (including the closed-source GPT-3.5-turbo), verifying the effectiveness and efficiency of our defense. Besides, we also preliminarily verify that BEAT can effectively defend against popular jailbreak attacks, as they can be regarded as "natural backdoors". Our source code is available at https://github.com/clearloveclearlove/BEAT.

Large language models (LLMs) can output sensitive information, which has emerged as a novel safety concern. Previous works focus on structured sensitive information (e.g. personal identifiable information). However, we notice that sensitive information can also be at semantic level, i.e. semantic sensitive information (SemSI). Particularly, *simple natural questions* can let state-of-the-art (SOTA) LLMs output SemSI. %which is hard to be detected compared with structured ones. Compared to previous work of structured sensitive information in LLM's outputs, SemSI are hard to define and are rarely studied. Therefore, we propose a novel and large-scale investigation on the existence of SemSI in SOTA LLMs induced by simple natural questions. First, we construct a comprehensive and labeled dataset of semantic sensitive information, SemSI-Set, by including three typical categories of SemSI. Then, we propose a large-scale benchmark, SemSI-Bench, to systematically evaluate semantic sensitive information in 25 SOTA LLMs. Our finding reveals that SemSI widely exists in SOTA LLMs' outputs by querying with simple natural questions. We open-source our project at https://semsi-project.github.io/.

Vision Language Model (VLM) Agents are stateful, autonomous entities capable of perceiving and interacting with their environments through vision and language. Multi-agent systems comprise specialized agents who collaborate to solve a (complex) task. A core security property is robustness, stating that the system maintains its integrity during adversarial attacks. Multi-agent systems lack robustness, as a successful exploit against one agent can spread and infect other agents to undermine the entire system’s integrity. We propose a defense Cowpox to provably enhance the robustness of a multi-agent system by a distributed mechanism that improves the recovery rate of agents by limiting the expected number of infections to other agents. The core idea is to generate and distribute a special cure sample that immunizes an agent against the attack before exposure. We demonstrate the effectiveness of Cowpox empirically and provide theoretical robustness guarantees.

Knowledge distillation (KD) is a vital technique for deploying deep neural networks (DNNs) on resource-constrained devices by transferring knowledge from large teacher models to lightweight student models. While teacher models from third-party platforms may undergo security verification (e.g., backdoor detection), we uncover a novel and critical threat: distillation-conditional backdoor attacks (DCBAs). DCBA injects dormant and undetectable backdoors into teacher models, which become activated in student models via the KD process, even with clean distillation datasets. While the direct extension of existing methods is ineffective for DCBA, we implement this attack by formulating it as a bilevel optimization problem and proposing a simple yet effective method (i.e., SCAR). Specifically, the inner optimization simulates the KD process by optimizing a surrogate student model, while the outer optimization leverages outputs from this surrogate to optimize the teacher model for implanting the conditional backdoor. Our SCAR addresses this complex optimization utilizing an implicit differentiation algorithm with a pre-optimized trigger injection function. Extensive experiments across diverse datasets, model architectures, and KD techniques validate the effectiveness of our SCAR and its resistance against existing backdoor detection, highlighting a significant yet previously overlooked vulnerability in the KD process. Our code is available at https://github.com/WhitolfChen/SCAR.

Multimodal large language models (MLLMs) remain vulnerable to transferable adversarial examples. While existing methods typically achieve targeted attacks by aligning global features—such as CLIP’s [CLS] token—between adversarial and target samples, they often overlook the rich local information encoded in patch tokens. This leads to suboptimal alignment and limited transferability, particularly for closed-source models. To address this limitation, we propose a targeted transferable adversarial attack method based on feature optimal alignment, called FOA-Attack, to improve adversarial transfer capability. Specifically, at the global level, we introduce a global feature loss based on cosine similarity to align the coarse-grained features of adversarial samples with those of target samples. At the local level, given the rich local representations within Transformers, we leverage clustering techniques to extract compact local patterns to alleviate redundant local features. We then formulate local feature alignment between adversarial and target samples as an optimal transport (OT) problem and propose a local clustering optimal transport loss to refine fine-grained feature alignment. Additionally, we propose a dynamic ensemble model weighting strategy to adaptively balance the influence of multiple models during adversarial example generation, thereby further improving transferability. Extensive experiments across various models demonstrate the superiority of the proposed method, outperforming state-of-the-art methods, especially in transferring to closed-source MLLMs.

Text-to-image (T2I) diffusion models have achieved impressive image generation quality and are increasingly fine-tuned for personalized applications. However, these models often inherit unsafe behaviors from toxic pretraining data, raising growing safety concerns. While recent safety-driven unlearning methods have made promising progress in suppressing model toxicity, they are found to be fragile to downstream fine-tuning, as we reveal that state-of-the-art methods largely fail to retain their effectiveness even when fine-tuned on entirely benign datasets. To mitigate this problem, in this paper, we propose ResAlign, a safety-driven unlearning framework with enhanced resilience against downstream fine-tuning. By modeling downstream fine-tuning as an implicit optimization problem with a Moreau envelope-based reformulation, ResAlign enables efficient gradient estimation to minimize the recovery of harmful behaviors. Additionally, a meta-learning strategy is proposed to simulate a diverse distribution of fine-tuning scenarios to improve generalization. Extensive experiments across a wide range of datasets, fine-tuning methods, and configurations demonstrate that ResAlign consistently outperforms prior unlearning approaches in retaining safety, while effectively preserving benign generation capability. Our code and pretrained models are publicly available at https://github.com/AntigoneRandy/ResAlign.

Multimodal Large Language Models (MLLMs) are increasingly deployed in fine-tuning-as-a-service (FTaaS) settings, where user-submitted datasets adapt general-purpose models to downstream tasks. This flexibility, however, introduces serious security risks, as malicious fine-tuning can implant backdoors into MLLMs with minimal effort. In this paper, we observe that backdoor triggers systematically disrupt cross-modal processing by causing abnormal attention concentration on non-semantic regions—a phenomenon we term **attention collapse**. Based on this insight, we propose **Believe Your Eyes (BYE)**, a data filtering framework that leverages attention entropy patterns as self-supervised signals to identify and filter backdoor samples. BYE operates via a three-stage pipeline: (1) extracting attention maps using the fine-tuned model, (2) computing entropy scores and profiling sensitive layers via bimodal separation, and (3) performing unsupervised clustering to remove suspicious samples. Unlike prior defenses, BYE equires no clean supervision, auxiliary labels, or model modifications. Extensive experiments across various datasets, models, and diverse trigger types validate BYE's effectiveness: it achieves near-zero attack success rates while maintaining clean-task performance, offering a robust and generalizable solution against backdoor threats in MLLMs.

Large language models (LLMs) have been widely applied for their remarkable capability of content generation. However, the practical use of open-source LLMs is hindered by high resource requirements, making deployment expensive and limiting widespread development. The collaborative inference is a promising solution for this problem, in which users collaborate by each hosting a subset of layers and transmitting intermediate activation. Many companies are building collaborative inference platforms to reduce LLM serving costs, leveraging users' underutilized GPUs. Despite widespread interest in collaborative inference within academia and industry, the privacy risks associated with LLM collaborative inference have not been well studied. This is largely because of the challenge posed by inverting LLM activation due to its strong non-linearity. In this paper, to validate the severity of privacy threats in LLM collaborative inference, we introduce the concept of prompt inversion attack (PIA), where a malicious participant intends to recover the input prompt through the activation transmitted by its previous participant. Specifically, we design a two-stage method to execute this attack. In the first stage, we optimize the input embedding with a constraint term derived from the LLM's embedding matrix to enforce the optimized embedding to be close to the ground truth. In the second stage, we accurately recover discrete tokens by incorporating activation calibration and semantic speculation. Extensive experiments show that our PIA method substantially outperforms existing baselines. For example, our method achieves an 88.4% token accuracy on the Skytrax dataset with the Llama-65B model when inverting the maximum number of transformer layers, while the best baseline method only achieves 22.8% accuracy. The results verify the effectiveness of our PIA attack and highlights its practical threat to LLM collaborative inference systems.

Text-to-image diffusion models are pushing the boundaries of what generative AI can achieve in our lives. Beyond their ability to generate general images, new personalization techniques have been proposed to customize the pretrained base models for crafting images with specific themes or styles. Such a lightweight solution, enabling AI practitioners and developers to easily build their own personalized models, also poses a new concern regarding whether the personalized models are trained from unauthorized data. A promising solution is to proactively enable data traceability in generative models, where data owners embed external coatings (e.g., image watermarks or backdoor triggers) onto the datasets before releasing. Later the models trained over such datasets will also learn the coatings and unconsciously reproduce them in the generated mimicries, which can be extracted and used as the data usage evidence. However, we identify the existing coatings cannot be effectively learned in personalization tasks, making the corresponding verification less reliable. In this paper, we introduce SIREN, a novel methodology to proactively trace unauthorized data usage in black-box personalized text-to-image diffusion models. Our approach optimizes the coating in a delicate way to be recognized by the model as a feature relevant to the personalization task, thus significantly improving its learnability. We also utilize a human perceptual-aware constraint, a hypersphere classification technique, and a hypothesis-testing-guided verification method to enhance the stealthiness and detection accuracy of the coating. The effectiveness of SIREN is verified through extensive experiments on a diverse set of benchmark datasets, models, and learning algorithms. SIREN is also effective in various real-world scenarios and evaluated against potential countermeasures. Our code is publicly available here.

Currently, deep neural networks (DNNs) are widely adopted in different applications. Despite its commercial values, training a well-performing DNN is resource-consuming. Accordingly, the well-trained model is valuable intellectual property for its owner. However, recent studies revealed the threats of model stealing, where the adversaries can obtain a function-similar copy of the victim model, even when they can only query the model. In this paper, we propose an effective and harmless model ownership verification (MOVE) to defend against different types of model stealing simultaneously, without introducing new security risks. In general, we conduct the ownership verification by verifying whether a suspicious model contains the knowledge of defender-specified external features. Specifically, we embed the external features by modifying a few training samples with style transfer. We then train a meta-classifier to determine whether a model is stolen from the victim. This approach is inspired by the understanding that the stolen models should contain the knowledge of features learned by the victim model. In particular, we develop our MOVE method under both glass-boxand closed-box settings and analyze its theoretical foundation to provide comprehensive model protection. Extensive experiments on benchmark datasets verify the effectiveness of our method and its resistance to potential adaptive attacks.

Currently, sample-specific backdoor attacks (SSBAs) are the most advanced and malicious methods since they can easily circumvent most of the current backdoor defenses. In this paper, we reveal that SSBAs are not sufficiently stealthy due to their poisoned-label nature, where users can discover anomalies if they check the image-label relationship. In particular, we demonstrate that it is ineffective to directly generalize existing SSBAs to their clean-label variants by poisoning samples solely from the target class. We reveal that it is primarily due to two reasons, including(1)the ‘antagonistic effects’ of ground-truth features and(2)the learning difficulty of sample-specific features. Accordingly, trigger-related features of existing SSBAs cannot be effectively learned under the clean-label setting due to their mild trigger intensity required for ensuring stealthiness. We argue that the intensity constraint of existing SSBAs is mostly because their trigger patterns are ‘content-irrelevant’ and therefore act as ‘noises’ for both humans and DNNs. Motivated by this understanding, we propose to exploit content-relevant features,$a.k.a.$(human-relied) attributes, as the trigger patterns to design clean-label SSBAs. This new attack paradigm is dubbed backdoor attack with attribute trigger (BAAT). Extensive experiments are conducted on benchmark datasets, which verify the effectiveness of our BAAT and its resistance to existing defenses.

Deep neural networks (DNNs) are susceptible to backdoor attacks, where adversaries poison datasets with adversary-specified triggers to implant hidden backdoors, enabling malicious manipulation of model predictions. Dataset purification serves as a proactive defense by removing malicious training samples to prevent backdoor injection at its source. We first reveal that the current advanced purification methods rely on a latent assumption that the backdoor connections between triggers and target labels in backdoor attacks are simpler to learn than the benign features. We demonstrate that this assumption, however, does not always hold, especially in all-to-all (A2A) and untargeted (UT) attacks. As a result, purification methods that analyze the separation between the poisoned and benign samples in the input-output space or the final hidden layer space are less effective. We observe that this separability is not confined to a single layer but varies across different hidden layers. Motivated by this understanding, we propose FLARE, a universal purification method to counter various backdoor attacks. FLARE aggregates abnormal activations from all hidden layers to construct representations for clustering. To enhance separation, FLARE develops an adaptive subspace selection algorithm to isolate the optimal space for dividing an entire dataset into two clusters. FLARE assesses the stability of each cluster and identifies the cluster with higher stability as poisoned. Extensive evaluations on benchmark datasets demonstrate the effectiveness of FLARE against 22 representative backdoor attacks, including all-to-one (A2O), all-to-all (A2A), and untargeted (UT) attacks, and its robustness to adaptive attacks.

Recently, point clouds have been widely used in computer vision, whereas their collection is time-consuming and expensive. As such, point cloud datasets are the valuable intellectual property of their owners and deserve protection. To detect and prevent unauthorized use of these datasets, especially for commercial or open-sourced ones that cannot be sold again or used commercially without permission, we intend to identify whether a suspicious third-party model is trained on our protected dataset under the black-box setting. We achieve this goal by designing ascalableclean-label backdoor-based dataset watermark for point clouds that ensures both effectiveness and stealthiness. Unlike existing clean-label watermark schemes, which were susceptible to the number of categories, our method can watermark samples from all classes instead of only from the target one. Accordingly, it can still preserve high effectiveness even on large-scale datasets with many classes. Specifically, we perturb selected point clouds with non-target categories in both shape-wise and point-wise manners before inserting trigger patterns without changing their labels. The features of perturbed samples are similar to those of benign samples from the target class. As such, models trained on the watermarked dataset will have a distinctive yet stealthy backdoor behavior,$i.e$., misclassifying samples from the target class whenever triggers appear, since the trained DNNs will treat the inserted trigger pattern as a signal to deny predicting the target label. We also design a hypothesis-test-guided dataset ownership verification based on the proposed watermark. Extensive experiments on benchmark datasets are conducted, verifying the effectiveness of our method and its resistance to potential removal methods. The codes are available athttps://github.com/weic0810/PointNCBW.

Backdoor attacks pose a significant threat to the security and reliability of deep learning models. To mitigate such attacks, one promising approach is to learn to extract features from the target model and use these features for backdoor detection. However, we discover that existing learning-based neural backdoor detection methods do not generalize well to new architectures not seen during the learning phase. In this paper, we analyze the root cause of this issue and propose a novel black-box neural backdoor detection method called ARCGEN. Our method aims to obtain architecture-invariant model features, i.e.,aligned features, for effective backdoor detection. Specifically, in contrast to existing methods directly using model outputs as model features, we introduce an additional alignment layer in the feature extraction function to further process these features. This reduces the direct influence of architecture information on the features. Then, we design two alignment losses to train the feature extraction function. These losses explicitly require that features from models with similar backdoor behaviors but different architectures are aligned at both the distribution and sample levels. With these techniques, our method demonstrates up to 42.5% improvements in detection performance (e.g., AUC) on unseen model architectures. This is based on a large-scale evaluation involving 16,896 models trained on diverse datasets, subjected to various backdoor attacks, and utilizing different model architectures. Our code is available at https://github.com/SeRAlab/ArcGen.

Although the widespread use of AI systems in today's world is growing, many current AI systems are found vulnerable due to hidden bias and missing information, especially in the most commonly used forecasting system. In this work, we explore the robustness and explainability of AI-based forecasting systems. We provide an in-depth analysis of the underlying causality involved in the effect prediction task and further establish a causal graph based on treatment, adjustment variable, confounder, and outcome. Correspondingly, we design a causal interventional prediction system (CIPS) based on a variational autoencoder and fully conditional specification of multiple imputations. Extensive results demonstrate the superiority of our system over state-of-the-art methods and show remarkable versatility and extensibility in practice.

Given the power of vision transformers, a new learning paradigm, pre-training and then prompting, makes it more efficient and effective to address downstream visual recog-nition tasks. In this paper, we identify a novel security threat towards such a paradigm from the perspective of back-door attacks. Specifically, an extra prompt token, called the switch token in this work, can turn the backdoor mode on, i.e., converting a benign model into a backdoored one. Once under the backdoor mode, a specific trigger can force the model to predict a target class. It poses a severe risk to the users of cloud API, since the malicious behavior can not be activated and detected under the benign mode, thus making the attack very stealthy. To attack a pre-trained model, our proposed attack, named SWARM, learns a trigger and prompt tokens including a switch token. They are optimized with the clean loss which encourages the model always be-haves normally even the trigger presents, and the backdoor loss that ensures the backdoor can be activated by the trig-ger when the switch is on. Besides, we utilize the cross-mode feature distillation to reduce the effect of the switch token on clean samples. The experiments on diverse vi-sual recognition tasks confirm the success of our switchable backdoor attack, i.e., achieving 95%+ attack success rate, and also being hard to be detected and removed. Our code is available at https://github.com/20000yshust/SWARM.

Recent studies revealed that using third-party models may lead to backdoor threats, where adversaries can maliciously manipulate model predictions based on backdoors implanted during model training. Arguably, backdoor trigger inversion (BTI), which generates trigger patterns of given benign samples for a backdoored model, is the most critical module for backdoor defenses used in these scenarios. With BTI, defenders can remove backdoors by fine-tuning based on generated poisoned samples with ground-truth labels or deactivate backdoors by removing trigger patterns during the inference process. However, we find that existing BTI methods suffer from relatively poor performance, $i.e.$, their generated triggers are significantly different from the ones used by the adversaries even in the feature space. We argue that it is mostly because existing methods require to 'extract' backdoor features at first, while this task is very difficult since defenders have no information ($e.g.$, trigger pattern or target label) about poisoned samples. In this paper, we explore BTI from another perspective where we decouple benign features instead of decoupling backdoor features directly. Specifically, our method consists of two main steps, including \textbf{(1)} decoupling benign features and \textbf{(2)} trigger inversion by minimizing the differences between benign samples and their generated poisoned version in decoupled benign features while maximizing the differences in remaining backdoor features. In particular, our method is more efficient since it doesn't need to `scan' all classes to speculate the target label, as required by existing BTI. We also exploit our BTI module to further design backdoor-removal and pre-processing-based defenses. Extensive experiments on benchmark datasets demonstrate that our defenses can reach state-of-the-art performances.

Saliency-based representation visualization (SRV) ($e.g.$, Grad-CAM) is one of the most classical and widely adopted explainable artificial intelligence (XAI) methods for its simplicity and efficiency. It can be used to interpret deep neural networks by locating saliency areas contributing the most to their predictions. However, it is difficult to automatically measure and evaluate the performance of SRV methods due to the lack of ground-truth salience areas of samples. In this paper, we revisit the backdoor-based SRV evaluation, which is currently the only feasible method to alleviate the previous problem. We first reveal its \emph{implementation limitations} and \emph{unreliable nature} due to the trigger generalization of existing backdoor watermarks. Given these findings, we propose a generalization-limited backdoor watermark (GLBW), based on which we design a more faithful XAI evaluation. Specifically, we formulate the training of watermarked DNNs as a min-max problem, where we find the `worst' potential trigger (with the highest attack effectiveness and differences from the ground-truth trigger) via inner maximization and minimize its effects and the loss over benign and poisoned samples via outer minimization in each iteration. In particular, we design an adaptive optimization method to find desired potential triggers in each inner maximization. Extensive experiments on benchmark datasets are conducted, verifying the effectiveness of our generalization-limited watermark. Our codes are available at \url{https://github.com/yamengxi/GLBW}.

Deep neural networks (DNNs) are vulnerable to backdoor attacks, where adversaries can maliciously trigger model misclassifications by implanting a hidden backdoor during model training. This paper proposes a simple yet effective input-level backdoor detection (dubbed IBD-PSC) as a ‘firewall’ to filter out malicious testing images. Our method is motivated by an intriguing phenomenon, i.e., parameter-oriented scaling consistency (PSC), where the prediction confidences of poisoned samples are significantly more consistent than those of benign ones when amplifying model parameters. In particular, we provide theoretical analysis to safeguard the foundations of the PSC phenomenon. We also design an adaptive method to select BN layers to scale up for effective detection. Extensive experiments are conducted on benchmark datasets, verifying the effectiveness and efficiency of our IBD-PSC method and its resistance to adaptive attacks. Codes are available at https://github.com/THUYimingLi/BackdoorBox.

Model quantization is a compression technique that converts a full-precision model to a more compact low-precision version for better storage. Despite the great success of quantization, recent studies revealed the feasibility of malicious exploiting model quantization via implanting quantization-conditioned backdoors (QCBs). These special backdoors remain dormant in full-precision models but are exposed upon quantization. Unfortunately, existing defenses have limited effects on mitigating QCBs. In this paper, we conduct an in-depth analysis of QCBs. We reveal an intriguing characteristic of QCBs, where activation of backdoor-related neurons on even benign samples enjoy a distribution drift after quantization, although this drift is more significant on poisoned samples. Motivated by this finding, we propose to purify the backdoor-exposed quantized model by aligning its layer-wise activation with its full-precision version. To further exploit the more pronounced activation drifts on poisoned samples, we design an additional module to layer-wisely approximate poisoned activation distribution based on batch normalization statistics of the full-precision model. Extensive experiments are conducted, verifying the effectiveness of our defense. Our code is publicly available.

High-quality public datasets significantly prompt the prosperity of deep neural networks (DNNs). Currently, dataset ownership verification (DOV), which consists of dataset watermarking and ownership verification, is the only feasible solution to protect their copyright by preventing unauthorized use. In this paper, we revisit existing DOV methods and find that they all mainly focused on the first stage by designing different types of dataset watermarks and directly exploiting watermarked samples as the verification samples for ownership verification. As such, their success relies on an underlying assumption that verification is a \emph{one-time} and \emph{privacy-preserving} process, which does not necessarily hold in practice. To alleviate this problem, we propose \emph{ZeroMark} to conduct ownership verification without disclosing dataset-specified watermarks. Our method is inspired by our empirical and theoretical findings of the intrinsic property of DNNs trained on the watermarked dataset. Specifically, ZeroMark first generates the closest boundary version of given benign samples and calculates their boundary gradients under the label-only black-box setting. After that, it examines whether the given suspicious method has been trained on the protected dataset by performing a hypothesis test, based on the cosine similarity measured on the boundary gradients and the watermark pattern. Extensive experiments on benchmark datasets verify the effectiveness of our ZeroMark and its resistance to potential adaptive attacks. The codes for reproducing our main experiments are publicly available at \href{https://github.com/JunfengGo/ZeroMark.git}{GitHub}.

Deep neural networks (DNNs) have been widely and successfully adopted and deployed in various applications of speech recognition. Recently, a few works revealed that these models are vulnerable to backdoor attacks, where the adversaries can implant malicious prediction behaviors into victim models by poisoning their training process. In this paper, we revisit poison-only backdoor attacks against speech recognition. We reveal that existing methods are not stealthy since their trigger patterns are perceptible to humans or machine detection. This limitation is mostly because their trigger patterns are simple noises or separable and distinctive clips. Motivated by these findings, we propose to exploit elements of sound (e.g., pitch and timbre) to design more stealthy yet effective poison-only backdoor attacks. Specifically, we insert a short-duration high-pitched signal as the trigger and increase the pitch of remaining audio clips to ‘mask’ it for designing stealthy pitch-based triggers. We manipulate timbre features of victim audio to design the stealthy timbre-based attack and design a voiceprint selection module to facilitate the multi-backdoor attack. Our attacks can generate more ‘natural’ poisoned samples and therefore are more stealthy. Extensive experiments are conducted on benchmark datasets, which verify the effectiveness of our attacks under different settings (e.g., all-to-one, all-to-all, clean-label, physical, and multi-backdoor settings) and their stealthiness. Our methods achieve attack success rates of over 95% in most cases and are nearly undetectable. The code for reproducing main experiments are available at https://github.com/HanboCai/BadSpeech_SoE.

Deep neural networks (DNNs) are vulnerable to backdoor attacks, where the adversary manipulates a small portion of training data such that the victim model predicts normally on the benign samples but classifies the triggered samples as the target class. The backdoor attack is an emerging yet threatening training-phase threat, leading to serious risks in DNN-based applications. In this paper, we revisit the trigger patterns of existing backdoor attacks. We reveal that they are either visible or not sparse and therefore are not stealthy enough. More importantly, it is not feasible to simply combine existing methods to design an effective sparse and invisible backdoor attack. To address this problem, we formulate the trigger generation as a bi-level optimization problem with sparsity and invisibility constraints and propose an effective method to solve it. The proposed method is dubbed sparse and invisible backdoor attack (SIBA). We conduct extensive experiments on benchmark datasets under different settings, which verify the effectiveness of our attack and its resistance to existing backdoor defenses. The codes for reproducing main experiments are available athttps://github.com/YinghuaGao/SIBA.

Backdoor attack intends to embed hidden backdoors into deep neural networks (DNNs), so that the attacked models perform well on benign samples, whereas their predictions will be maliciously changed if the hidden backdoor is activated by attacker-specified triggers. This threat could happen when the training process is not fully controlled, such as training on third-party datasets or adopting third-party models, which poses a new and realistic threat. Although backdoor learning is an emerging and rapidly growing research area, there is still no comprehensive and timely review of it. In this article, we present the first comprehensive survey of this realm. We summarize and categorize existing backdoor attacks and defenses based on their characteristics, and provide a unified framework for analyzing poisoning-based backdoor attacks. Besides, we also analyze the relation between backdoor attacks and relevant fields (i.e., adversarial attacks and data poisoning), and summarize widely adopted benchmark datasets. Finally, we briefly outline certain future research directions relying upon reviewed works. A curated list of backdoor-related resources is also available at https://github.com/THUYimingLi/backdoor-learning-resources.

Recent studies have demonstrated that existing deep neural networks (DNNs) on 3D point clouds are vulnerable to adversarial examples, especially under the white-box settings where the adversaries have access to model parameters. However, adversarial 3D point clouds generated by existing white-box methods have limited transferability across different DNN architectures. They have only minor threats in real-world scenarios under the black-box settings where the adversaries can only query the deployed victim model. In this paper, we revisit the transferability of adversarial 3D point clouds. We observe that an adversarial perturbation can be randomly factorized into two sub-perturbations, which are also likely to be adversarial perturbations. It motivates us to consider the effects of the perturbation and its sub-perturbations simultaneously to increase the transferability for sub-perturbations also contain helpful information. In this paper, we propose a simple yet effective attack method to generate more transferable adversarial 3D point clouds. Specifically, rather than simply optimizing the loss of perturbation alone, we combine it with its random factorization. We conduct experiments on benchmark dataset, verifying our method's effectiveness in increasing transferability while preserving high efficiency.

Recent studies revealed that deep neural networks (DNNs) are exposed to backdoor threats when training with third-party resources (such as training samples or backbones). The back-doored model has promising performance in predicting benign samples, whereas its predictions can be maliciously manipulated by adversaries based on activating its backdoors with pre-defined trigger patterns. Currently, most of the existing backdoor attacks were conducted on the image classification under the targeted manner. In this paper, we reveal that these threats could also happen in object detection, posing threatening risks to many mission-critical applications (e.g., pedestrian detection and intelligent surveillance systems). Specifically, we design a simple yet effective poison-only backdoor attack in an untargeted manner, based on task characteristics. We show that, once the backdoor is embedded into the target model by our attack, it can trick the model to lose detection of any object stamped with our trigger patterns. We conduct extensive experiments on the benchmark dataset, showing its effectiveness in both digital and physical-world settings and its resistance to potential defenses.

Deep neural networks (DNNs) are vulnerable to backdoor attacks. The backdoor adversaries intend to maliciously control the predictions of attacked DNNs by injecting hidden backdoors that can be activated by adversary-specified trigger patterns during the training process. One recent research revealed that most of the existing attacks failed in the real physical world since the trigger contained in the digitized test samples may be different from that of the one used for training. Accordingly, users can adopt spatial transformations as the image pre-processing to deactivate hidden backdoors. In this paper, we explore the previous findings from another side. We exploit classical spatial transformations (i.e., rotation and translation) with the specific parameter as trigger patterns to design a simple yet effective poisoning-based backdoor attack. For example, only images rotated to a particular angle can activate the embedded backdoor of attacked DNNs. Extensive experiments are conducted, verifying the effectiveness of our attack under both digital and physical settings and its resistance to existing backdoor defenses.

Recent studies have demonstrated that deep neural networks (DNNs) are vulnerable to backdoor attacks during the training process. Specifically, the adversaries intend to embed hidden backdoors in DNNs so that malicious model predictions can be activated through pre-defined trigger patterns. In this paper, we explore the backdoor mechanism from the angle of the model structure. We select the skip connection for discussions, inspired by the understanding that it helps the learning of model ‘shortcuts’ where backdoor triggers are usually easier to be learned. Specifically, we demonstrate that the attack success rate (ASR) decreases significantly when reducing the outputs of some key skip connections. Based on this observation, we design a simple yet effective backdoor removal method by suppressing the skip connections in critical layers selected by our method. We also implement fine-tuning on these layers to recover high benign accuracy and to further reduce ASR. Extensive experiments on benchmark datasets verify the effectiveness of our method.

Deep neural networks (DNNs) are widely deployed on real-world devices. Concerns regarding their security have gained great attention from researchers. Recently, a new weight modification attack called bit flip attack (BFA) was proposed, which exploits memory fault inject techniques such as row hammer to attack quantized models in the deployment stage. With only a few bit flips, the target model can be rendered useless as a random guesser or even be implanted with malicious functionalities. In this work, we seek to further reduce the number of bit flips. We propose a training-assisted bit flip attack, in which the adversary is involved in the training stage to build a high-risk model to release. This high-risk model, obtained coupled with a corresponding malicious model, behaves normally and can escape various detection methods. The results on benchmark datasets show that an adversary can easily convert this high-risk but normal model to a malicious one on victim’s side by flipping only one critical bit on average in the deployment stage. Moreover, our attack still poses a significant threat even when defenses are employed. The codes for reproducing main experiments are available at https://github.com/jianshuod/TBA.

Deep neural networks are valuable assets considering their commercial benefits and huge demands for costly annotation and computation resources. To protect the copyright of DNNs, backdoor-based ownership verification becomes popular recently, in which the model owner can watermark the model by embedding a specific backdoor behavior before releasing it. The defenders (usually the model owners) can identify whether a suspicious third-party model is "stolen" from them based on the presence of the behavior. Unfortunately, these watermarks are proven to be vulnerable to removal attacks even like fine-tuning. To further explore this vulnerability, we investigate the parameter space and find there exist many watermark-removed models in the vicinity of the watermarked one, which may be easily used by removal attacks. Inspired by this finding, we propose a mini-max formulation to find these watermark-removed models and recover their watermark behavior. Extensive experiments demonstrate that our method improves the robustness of the model watermarking against parametric changes and numerous watermark-removal attacks. The codes for reproducing our main experiments are available at https://github.com/GuanhaoGan/robust-model-watermarking.

The prosperity of deep neural networks (DNNs) is largely benefited from open-source datasets, based on which users can evaluate and improve their methods. In this paper, we revisit backdoor-based dataset ownership verification (DOV), which is currently the only feasible approach to protect the copyright of open-source datasets. We reveal that these methods are fundamentally harmful given that they could introduce malicious misclassification behaviors to watermarked DNNs by the adversaries. In this paper, we design DOV from another perspective by making watermarked models (trained on the protected dataset) correctly classify some `hard' samples that will be misclassified by the benign model. Our method is inspired by the generalization property of DNNs, where we find a \emph{hardly-generalized domain} for the original dataset (as its \emph{domain watermark}). It can be easily learned with the protected dataset containing modified samples. Specifically, we formulate the domain generation as a bi-level optimization and propose to optimize a set of visually-indistinguishable clean-label modified data with similar effects to domain-watermarked samples from the hardly-generalized domain to ensure watermark stealthiness. We also design a hypothesis-test-guided ownership verification via our domain watermark and provide the theoretical analyses of our method. Extensive experiments on three benchmark datasets are conducted, which verify the effectiveness of our method and its resistance to potential adaptive methods.

Deep learning, especially deep neural networks (DNNs), has been widely and successfully adopted in many critical applications for its high effectiveness and efficiency. The rapid development of DNNs has benefited from the existence of some high-quality datasets (e.g., ImageNet), which allow researchers and developers to easily verify the performance of their methods. Currently, almost all existing released datasets require that they can only be adopted for academic or educational purposes rather than commercial purposes without permission. However, there is still no good way to ensure that. In this paper, we formulate the protection of released datasets as verifying whether they are adopted for training a (suspicious) third-party model, where defenders can only query the model while having no information about its parameters and training details. Based on this formulation, we propose to embed external patterns via backdoor watermarking for the ownership verification to protect them. Our method contains two main parts, including dataset watermarking and dataset verification. Specifically, we exploit poison-only backdoor attacks (e.g., BadNets) for dataset watermarking and design a hypothesis-test-guided method for dataset verification. We also provide some theoretical analyses of our methods. Experiments on multiple benchmark datasets of different tasks are conducted, which verify the effectiveness of our method. The code for reproducing main experiments is available at https://github.com/THUYimingLi/DVBW.

Obtaining a well-trained model involves expensive data collection and training procedures, therefore the model is a valuable intellectual property. Recent studies revealed that adversaries can `steal' deployed models even when they have no training samples and can not get access to the model parameters or structures. Currently, there were some defense methods to alleviate this threat, mostly by increasing the cost of model stealing. In this paper, we explore the defense from another angle by verifying whether a suspicious model contains the knowledge of defender-specified external features. Specifically, we embed the external features by tempering a few training samples with style transfer. We then train a meta-classifier to determine whether a model is stolen from the victim. This approach is inspired by the understanding that the stolen models should contain the knowledge of features learned by the victim model. We examine our method on both CIFAR-10 and ImageNet datasets. Experimental results demonstrate that our method is effective in detecting different types of model stealing simultaneously, even if the stolen model is obtained via a multi-stage stealing process. The codes for reproducing main results are available at Github (https://github.com/zlh-thu/StealingVerification).

Image representation is critical for many visual tasks. Instead of representing images discretely with 2D arrays of pixels, a recent study, namely local implicit image function (LIIF), denotes images as a continuous function where pixel values are expansion by using the corresponding coordinates as inputs. Due to its continuous nature, LIIF can be adopted for arbitrary-scale image super-resolution tasks, resulting in a single effective and efficient model for various up-scaling factors. However, LIIF often suffers from structural distortions and ringing artifacts around edges, mostly because all pixels share the same model, thus ignoring the local properties of the image. In this paper, we propose a novel adaptive local image function (A-LIIF) to alleviate this problem. Specifically, our A-LIIF consists of two main components: an encoder and a expansion network. The former captures cross-scale image features, while the latter models the continuous up-scaling function by a weighted combination of multiple local implicit image functions. Accordingly, our A-LIIF can reconstruct the high-frequency textures and structures more accurately. Experiments on multiple benchmark datasets verify the effectiveness of our method. Our codes are available at https://github.com/LeeHW-THU/A-LIIF.

Deep neural networks (DNNs) have demonstrated their superiority in practice. Arguably, the rapid development of DNNs is largely benefited from high-quality (open-sourced) datasets, based on which researchers and developers can easily evaluate and improve their learning methods. Since the data collection is usually time-consuming or even expensive, how to protect their copyrights is of great significance and worth further exploration. In this paper, we revisit dataset ownership verification. We find that existing verification methods introduced new security risks in DNNs trained on the protected dataset, due to the targeted nature of poison-only backdoor watermarks. To alleviate this problem, in this work, we explore the untargeted backdoor watermarking scheme, where the abnormal model behaviors are not deterministic. Specifically, we introduce two dispersibilities and prove their correlation, based on which we design the untargeted backdoor watermark under both poisoned-label and clean-label settings. We also discuss how to use the proposed untargeted backdoor watermark for dataset ownership verification. Experiments on benchmark datasets verify the effectiveness of our methods and their resistance to existing backdoor defenses.

Federated learning (FL) enables data owners to train a global model with shared gradients while keeping private training data locally. However, recent research demonstrated that the adversary may infer private training data of clients from the exchanged local gradients, e.g., having deep leakage from gradients (DLGs). Many existing privacy-preserving approaches take usage of differential privacy (DP) to guarantee privacy. Nevertheless, the widely used privacy budget of DP (e.g., evenly distribution) leads to a sharp decline of model accuracy. To improve the model accuracy, some schemes only consider allocating the privacy budget to the fully connected layers. However, we reveal that the adversary may still reconstruct the private training data by adopting the DLG attack with the gradients of convolutional layers. In this article, we propose a fine-grained DP federated learning (DPFL) scheme, which guarantees privacy and remains high model performance simultaneously. Specifically, inspired by the methods that measure the importance of layers in deep learning, we propose a fine-grained method to allocate noise according to the importance value of layers in order to remain high model performance. Besides, we combine an active client selection strategy with DPFL and perform fine-tuning with a public data set on the server to further ensure the model performance. We evaluate DPFL under both independent and identically distributed (i.i.d) and non-i.i.d data settings to show that our method can achieve similar accuracy as the plain FL (e.g., FedAvg). We also demonstrate that our DPFL can resist the DLG attack to verify its privacy guarantee.

Privacy protection is an important research area, which is especially critical in this big data era. To a large extent, the privacy of visual classification data is mainly in the mapping between the image and its corresponding label, since this relation provides a great amount of information and can be used in other scenarios. In this paper, we propose the mapping distortion based protection (MDP) and its augmentation-based extension (AugMDP) to protect the data privacy by modifying the original dataset. In the modified dataset generated by MDP, the image and its label are not consistent (e.g., a cat-like image is labeled as the dog), whereas the DNNs trained on it can still achieve good performance on benign testing set. As such, this method can protect privacy when the dataset is leaked. Extensive experiments are conducted, which verify the effectiveness and feasibility of our method. The code for reproducing main results is available at https://github.com/PerdonLiu/Visual-Privacy-Protection-via-Mapping-Distortion.

k-means algorithm is one of the most classical clustering methods, which has been widely and successfully used in signal processing. However, due to the thin-tailed property of the Gaussian distribution, k-means algorithm suffers from relatively poor performance on the dataset containing heavy-tailed data or outliers. Besides, standard k-means algorithm also has relatively weak stability, i.e. its results have a large variance, which reduces its credibility. In this paper, we propose a robust and stable k-means variant, dubbed the t-k-means, as well as its fast version to alleviate those problems. Theoretically, we derive the t-k-means and analyze its robustness and stability from the aspect of the loss function and the expression of the clustering center, respectively. Extensive experiments are also conducted, which verify the effectiveness and efficiency of the proposed method. The code for reproducing main results is available at https://github.com/THUYimingLi/t-k-means.

Speaker verification has been widely and successfully adopted in many mission-critical areas for user identification. The training of speaker verification requires a large amount of data, therefore users usually need to adopt third-party data (e.g., data from the Internet or third-party data company). This raises the question of whether adopting untrusted third-party data can pose a security threat. In this paper, we demonstrate that it is possible to inject the hidden backdoor for infecting speaker verification models by poisoning the training data. Specifically, we design a clustering-based attack scheme where poisoned samples from different clusters will contain different triggers (i.e., pre-defined utterances), based on our understanding of verification tasks. The infected models behave normally on benign samples, while attacker-specified unenrolled triggers will successfully pass the verification even if the attacker has no information about the enrolled speaker. We also demonstrate that existing back-door attacks cannot be directly adopted in attacking speaker verification. Our approach not only provides a new perspective for designing novel attacks, but also serves as a strong baseline for improving the robustness of verification methods. The code for reproducing main results is available at https://github.com/zhaitongqing233/Backdoor-attack-against-speaker-verification.

Recently, backdoor attacks pose a new security threat to the training process of deep neural networks (DNNs). Attackers intend to inject hidden backdoors into DNNs, such that the attacked model performs well on benign samples, whereas its prediction will be maliciously changed if hidden backdoors are activated by the attacker-defined trigger. Existing backdoor attacks usually adopt the setting that triggers are sample-agnostic, i.e., different poisoned samples contain the same trigger, resulting in that the attacks could be easily mitigated by current backdoor defenses. In this work, we explore a novel attack paradigm, where backdoor triggers are sample-specific. In our attack, we only need to modify certain training samples with invisible perturbation, while not need to manipulate other training components (e.g., training loss, and model structure) as required in many existing attacks. Specifically, inspired by the recent advance in DNN-based image steganography, we generate sample-specific invisible additive noises as backdoor triggers by encoding an attacker-specified string into benign images through an encoder-decoder network. The mapping from the string to the target label will be generated when DNNs are trained on the poisoned dataset. Extensive experiments on benchmark datasets verify the effectiveness of our method in attacking models with or without defenses. The code will be available at https://github.com/yuezunli/ISSBA.

Adversarial defense is a popular and important research area. Due to its intrinsic mechanism, one of the most straightforward and effective ways of defending attacks is to analyze the property of loss surface in the input space. In this paper, we define the local flatness of the loss surface as the maximum value of the chosen norm of the gradient regarding to the input within a neighborhood centered on the benign sample, and discuss the relationship between the local flatness and adversarial vulnerability. Based on the analysis, we propose a novel defense approach via regularizing the local flatness, dubbed local flatness regularization (LFR). We also demonstrate the effectiveness of the proposed method from other perspectives, such as human visual mechanism, and analyze the relationship between LFR and other related methods theoretically. Experiments are conducted to verify our theory and demonstrate the superiority of the proposed method.

Despite being one of the most popular nonparametric approaches, Gaussian process regression (GPR) suffers from O(n3) computational burden and the computation is infeasible for large-scale scenarios. To reduce the computational complexity, many Shannon-mutual-information-based aggregation methods were proposed, whereas these methods can not effectively identify the importance of experts in some cases. To address this problem, we generalize the traditional mutual information-based methods (GPoE, RBCM, GRBCM) based on Tsallis mutual information. Accordingly, the generated weight distribution is more sparse tending to focus on those experts with good performance. To obtain adaptive and data-dependent entropic-index in Tsallis entropy, we propose three heuristic algorithms to solve our model. Extensive experiments show that, the proposed method can improve the prediction of both the mean and variance, and the improvement of variance prediction is significant in many cases.

From the mutual empowerment of two high-speed development technologies: artificial intelligence and edge computing, we propose a tailored Edge Intelligent Video Surveillance (EIVS) system. It is a scalable edge computing architecture and uses multitask deep learning for relevant computer vision tasks. Due to the potential application of different surveillance devices are widely different, we adopt a smart IoT module to normalize the video data of different cameras, thus the EIVS system can conveniently found proper data for a specific task. In addition, the deep learning models can be deployed at every EIVS nodes, to make computer vision tasks on the normalized data. Meanwhile, due to the training and deploying of deep learning model are usually separated, for the related tasks in the same scenario, we propose to collaboratively train the depth learning models in a multitask paradigm on the cloud server. The simulation results on the publicly available datasets show that the system continuously supports intelligent monitoring tasks, has good scalability, and can improve performance through multitask learning.