Yingjie Lao

Khoa D. Doan, Yingjie Lao, Ping Li · baidu

In recent years, machine learning models have been shown to be vulnerable to backdoor attacks. Under such attacks, an adversary embeds a stealthy backdoor into the trained model such that the compromised models will behave normally on clean inputs but will misclassify according to the adversary's control on maliciously constructed input with a trigger. While these existing attacks are very effective, the adversary's capability is limited: given an input, these attacks can only cause the model to misclassify toward a single pre-defined or target class. In contrast, this paper exploits a novel backdoor attack with a much more powerful payload, denoted as Marksman, where the adversary can arbitrarily choose which target class the model will misclassify given any input during inference. To achieve this goal, we propose to represent the trigger function as a class-conditional generative model and to inject the backdoor in a constrained optimization framework, where the trigger function learns to generate an optimal trigger pattern to attack any target class at will while simultaneously embedding this generative backdoor into the trained model. Given the learned trigger-generation function, during inference, the adversary can specify an arbitrary backdoor attack target class, and an appropriate trigger causing the model to classify toward this target class is created accordingly. We show empirically that the proposed framework achieves high attack performance while preserving the clean-data performance in several benchmark datasets, including MNIST, CIFAR10, GTSRB, and TinyImageNet. The proposed Marksman backdoor attack can also easily bypass existing backdoor defenses that were originally designed against backdoor attacks with a single target class. Our work takes another significant step toward understanding the extensive risks of backdoor attacks in practice.

Khoa D. Doan, Yingjie Lao, Peng Yang et al. · baidu

Vision Transformers (ViTs) have a radically different architecture with significantly less inductive bias than Convolutional Neural Networks. Along with the improvement in performance, security and robustness of ViTs are also of great importance to study. In contrast to many recent works that exploit the robustness of ViTs against adversarial examples, this paper investigates a representative causative attack, i.e., backdoor. We first examine the vulnerability of ViTs against various backdoor attacks and find that ViTs are also quite vulnerable to existing attacks. However, we observe that the clean-data accuracy and backdoor attack success rate of ViTs respond distinctively to patch transformations before the positional encoding. Then, based on this finding, we propose an effective method for ViTs to defend both patch-based and blending-based trigger backdoor attacks via patch processing. The performances are evaluated on several benchmark datasets, including CIFAR10, GTSRB, and TinyImageNet, which show the proposed novel defense is very successful in mitigating backdoor attacks for ViTs. To the best of our knowledge, this paper presents the first defensive strategy that utilizes a unique characteristic of ViTs against backdoor attacks. The paper will appear in the Proceedings of the AAAI'23 Conference. This work was initially submitted in November 2021 to CVPR'22, then it was re-submitted to ECCV'22. The paper was made public in June 2022. The authors sincerely thank all the referees from the Program Committees of CVPR'22, ECCV'22, and AAAI'23.

Quang H. Nguyen, Yingjie Lao, Tung Pham et al. · baidu

Recent works have shown that deep neural networks are vulnerable to adversarial examples that find samples close to the original image but can make the model misclassify. Even with access only to the model's output, an attacker can employ black-box attacks to generate such adversarial examples. In this work, we propose a simple and lightweight defense against black-box attacks by adding random noise to hidden features at intermediate layers of the model at inference time. Our theoretical analysis confirms that this method effectively enhances the model's resilience against both score-based and decision-based black-box attacks. Importantly, our defense does not necessitate adversarial training and has minimal impact on accuracy, rendering it applicable to any pre-trained model. Our analysis also reveals the significance of selectively adding noise to different parts of the model based on the gradient of the adversarial objective function, which can be varied during the attack. We demonstrate the robustness of our defense against multiple black-box attacks through extensive empirical experiments involving diverse models with various architectures.

Pang Liu, Yingjie Lao

Universal adversarial attacks on aligned multimodal large language models are increasingly reported with attack success rates in the 60-80% range, suggesting the visual modality is highly vulnerable to imperceptible perturbations as a prompt-injection channel. We argue that this number conflates two distinct events: (i) the model's output was perturbed (Influence), and (ii) the attacker's chosen target concept was actually emitted (Precise Injection). We compose two existing techniques -- Universal Adversarial Attack and AnyAttack -- under an $L_{inf}$ budget of 16/255, and we add a dual-axis evaluation: a deterministic Ratcliff-Obershelp drift score for Influence (programmatic baseline) plus a 4-tier ordinal categorical none/weak/partial/confirmed for Precise Injection. The judge is DeepSeek-V4-Pro in thinking mode, calibrated against Claude Opus 4.7 with Cohen's $κ$ = 0.77 on the injection axis (substantial agreement); the entire 4475-entry SHA-256 input cache ships with the dataset so reviewers can re-derive paper numbers bit-exact without an API key. Across 6615 pairs over four open VLMs, seven attack prompts, and seven test images, the two axes diverge by roughly 90$\times$: 66.4% of pairs are programmatically disturbed (LLM-judged 46.6% at the substantial-or-complete tier), but only 0.756% (50/6615) reach any non-none injection tier and only 0.030% (2/6615) verbatim. The few injections that do land cluster on screenshot- or document-style carriers whose semantics already invite text transcription. BLIP-2 shows \emph{zero detectable drift} at $L_{inf}$ = 16/255 across all 2205 pairs even when used as a Stage-1 surrogate. We release the full dataset -- 21 universal images, 147 adversarial photos, 6,615 response pairs, the v3 dual-axis judge results, and the cache at huggingface.co/datasets/jeffliulab/visinject.

Faysal Hossain Shezan, Yingjie Lao, Minlong Peng et al.

The recent privacy leakage incidences and the more strict policy regulations demand a much higher standard of compliance for companies and mobile apps. However, such obligations also impose significant challenges on app developers for complying with these regulations that contain various perspectives, activities, and roles, especially for small companies and developers who are less experienced in this matter or with limited resources. To address these hurdles, we develop an automatic tool, NL2GDPR, which can generate policies from natural language descriptions from the developer while also ensuring the app's functionalities are compliant with General Data Protection Regulation (GDPR). NL2GDPR is developed by leveraging an information extraction tool, OIA (Open Information Annotation), developed by Baidu Cognitive Computing Lab. At the core, NL2GDPR is a privacy-centric information extraction model, appended with a GDPR policy finder and a policy generator. We perform a comprehensive study to grasp the challenges in extracting privacy-centric information and generating privacy policies, while exploiting optimizations for this specific task. With NL2GDPR, we can achieve 92.9%, 95.2%, and 98.4% accuracy in correctly identifying GDPR policies related to personal data storage, process, and share types, respectively. To the best of our knowledge, NL2GDPR is the first tool that allows a developer to automatically generate GDPR compliant policies, with only the need of entering the natural language for describing the app features. Note that other non-GDPR-related features might be integrated with the generated features to build a complex app.

Jun Woo Chung, Yingjie Lao, Weijie Zhao

Gradient Boosting Decision Trees (GBDTs) are widely used in industry and academia for their high accuracy and efficiency, particularly on structured data. However, watermarking GBDT models remains underexplored compared to neural networks. In this work, we present the first robust watermarking framework tailored to GBDT models, utilizing in-place fine-tuning to embed imperceptible and resilient watermarks. We propose four embedding strategies, each designed to minimize impact on model accuracy while ensuring watermark robustness. Through experiments across diverse datasets, we demonstrate that our methods achieve high watermark embedding rates, low accuracy degradation, and strong resistance to post-deployment fine-tuning.

Bingyin Zhao, Zhiding Yu, Shiyi Lan et al.

Recent studies indicate that Vision Transformers (ViTs) are robust against out-of-distribution scenarios. In particular, the Fully Attentional Network (FAN) - a family of ViT backbones, has achieved state-of-the-art robustness. In this paper, we revisit the FAN models and improve their pre-training with a self-emerging token labeling (STL) framework. Our method contains a two-stage training framework. Specifically, we first train a FAN token labeler (FAN-TL) to generate semantically meaningful patch token labels, followed by a FAN student model training stage that uses both the token labels and the original class label. With the proposed STL framework, our best model based on FAN-L-Hybrid (77.3M parameters) achieves 84.8% Top-1 accuracy and 42.1% mCE on ImageNet-1K and ImageNet-C, and sets a new state-of-the-art for ImageNet-A (46.1%) and ImageNet-R (56.6%) without using extra data, outperforming the original FAN counterpart by significant margins. The proposed framework also demonstrates significantly enhanced performance on downstream tasks such as semantic segmentation, with up to 1.7% improvement in robustness over the counterpart model. Code is available at https://github.com/NVlabs/STL.

Rui Chu, Bingyin Zhao, Thanh Quoc Hung Le et al.

Large language models (LLMs) have achieved unprecedented success due to their exceptional generative capabilities. However, because they depend on knowledge encapsulated from training corpora, they may produce hallucinations, stereotypes, and socially biased content. In particular, LLMs are prone to prejudiced responses involving race, gender, and age, which are collectively referred to as social biases. Prior studies have used fine-tuning and prompt engineering to mitigate such biases in LLMs, but these methods require additional training resources or domain knowledge to design the framework. Moreover, they may degrade the original capabilities of LLMs and often overlook the need for dynamic debiasing contexts for fairer inference. In this paper, we propose DebiasRAG, a novel tuning-free and dynamic query-specific debiasing framework based on retrieval-augmented generation (RAG). DebiasRAG improves fairness while preserving the intrinsic properties of LLMs, such as representation ability. DebiasRAG consists of three stages: (1) query-specific debiasing candidate generation; (2) context candidate pool construction; and (3) gradient-updated debiasing-guided context piece reranking. First, DebiasRAG leverages self-diagnosed bias contexts relevant to the query through regular retrieval, where the bias contexts are prepared offline by the DebiasRAG provider. Given the query-specific bias contexts, DebiasRAG reversely produces debiasing contexts, which are provided as additional fairness constraints for LLM outputs. Second, a regular RAG retrieval process produces query-related contexts from the regular RAG document database, such as a chunked Wikipedia dataset.

Nicole Meng, Zheyuan Liu, Meng Jiang et al.

Recent advances in 3D generative editing, particularly pipelines based on 3D Gaussian Splatting (3DGS), have achieved high-fidelity, multi-view-consistent scene manipulation from text prompts. However, we find that these pipelines also introduce new safety risks when unsafe prompts produce edits that are propagated and optimized across views. In this work, we study unsafe generation in 3D editing pipelines and show that such behavior can lead to coherent, undesirable Not-Safe-For-Work (NSFW) content in the final 3D representation. To address this, we propose 3DEditSafe, a safety-regularized 3D editing framework that constrains unsafe semantic propagation during optimization. 3DEditSafe combines generation-stage safety guidance with rendered-view 3D safety regularization, safe semantic projection, residue suppression, and mask-aware preservation to steer optimization away from unsafe editing directions. We evaluate our approach on EditSplat scenes using an object-compatible unsafe prompt benchmark and show that 2D safety guidance alone is not consistently sufficient to prevent unsafe 3D edits. 3DEditSafe reduces unsafe semantic alignment and view-level attack success rates, while revealing a safety-quality tradeoff in which stronger unsafe suppression can introduce artifacts or reduce unsafe-prompt fidelity. To our knowledge, this work is the first attempt to study and defend against unsafe generation in text-driven 3D editing pipelines, highlighting the need for safety mechanisms that operate directly on optimized 3D representations.

Kiran Thorat, Nicole Meng, Mostafa Karami et al.

IR drop analysis is essential in physical chip design to ensure the power integrity of on-chip power delivery networks. Traditional Electronic Design Automation (EDA) tools have become slow and expensive as transistor density scales. Recent works have introduced machine learning (ML)-based methods that formulate IR drop analysis as an image prediction problem. These existing ML approaches fail to capture both local and long-range dependencies and ignore crucial geometrical and topological information from physical layouts and logical connectivity. To address these limitations, we propose GIF, a Generative IR drop Framework that uses both geometrical and topological information to generate IR drop images. GIF fuses image and graph features to guide a conditional diffusion process, producing high-quality IR drop images. For instance, On the CircuitNet-N28 dataset, GIF achieves 0.78 SSIM, 0.95 Pearson correlation, 21.77 PSNR, and 0.026 NMAE, outperforming prior methods. These results demonstrate that our framework, using diffusion based multimodal conditioning, reliably generates high quality IR drop images. This shows that IR drop analysis can effectively leverage recent advances in generative modeling when geometric layout features and logical circuit topology are jointly modeled. By combining geometry aware spatial features with logical graph representations, GIF enables IR drop analysis to benefit from recent advances in generative modeling for structured image generation.

Huawei Lin, Yingjie Lao, Tong Geng et al.

Large Language Models (LLMs) are vulnerable to attacks like prompt injection, backdoor attacks, and adversarial attacks, which manipulate prompts or models to generate harmful outputs. In this paper, departing from traditional deep learning attack paradigms, we explore their intrinsic relationship and collectively term them Prompt Trigger Attacks (PTA). This raises a key question: Can we determine if a prompt is benign or poisoned? To address this, we propose UniGuardian, the first unified defense mechanism designed to detect prompt injection, backdoor attacks, and adversarial attacks in LLMs. Additionally, we introduce a single-forward strategy to optimize the detection pipeline, enabling simultaneous attack detection and text generation within a single forward pass. Our experiments confirm that UniGuardian accurately and efficiently identifies malicious prompts in LLMs.

Huawei Lin, Yingjie Lao, Weijie Zhao

Identifying the training data samples that most influence a generated image is a critical task in understanding diffusion models (DMs), yet existing influence estimation methods are constrained to small-scale or LoRA-tuned models due to computational limitations. To address this challenge, we propose DMin (Diffusion Model influence), a scalable framework for estimating the influence of each training data sample on a given generated image. To the best of our knowledge, it is the first method capable of influence estimation for DMs with billions of parameters. Leveraging efficient gradient compression, DMin reduces storage requirements from hundreds of TBs to mere MBs or even KBs, and retrieves the top-k most influential training samples in under 1 second, all while maintaining performance. Our empirical results demonstrate DMin is both effective in identifying influential training samples and efficient in terms of computational and storage requirements.

Yuning Han, Bingyin Zhao, Rui Chu et al.

Recent studies show that diffusion models (DMs) are vulnerable to backdoor attacks. Existing backdoor attacks impose unconcealed triggers (e.g., a gray box and eyeglasses) that contain evident patterns, rendering remarkable attack effects yet easy detection upon human inspection and defensive algorithms. While it is possible to improve stealthiness by reducing the strength of the backdoor, doing so can significantly compromise its generality and effectiveness. In this paper, we propose UIBDiffusion, the universal imperceptible backdoor attack for diffusion models, which allows us to achieve superior attack and generation performance while evading state-of-the-art defenses. We propose a novel trigger generation approach based on universal adversarial perturbations (UAPs) and reveal that such perturbations, which are initially devised for fooling pre-trained discriminative models, can be adapted as potent imperceptible backdoor triggers for DMs. We evaluate UIBDiffusion on multiple types of DMs with different kinds of samplers across various datasets and targets. Experimental results demonstrate that UIBDiffusion brings three advantages: 1) Universality, the imperceptible trigger is universal (i.e., image and model agnostic) where a single trigger is effective to any images and all diffusion models with different samplers; 2) Utility, it achieves comparable generation quality (e.g., FID) and even better attack success rate (i.e., ASR) at low poison rates compared to the prior works; and 3) Undetectability, UIBDiffusion is plausible to human perception and can bypass Elijah and TERD, the SOTA defenses against backdoors for DMs. We will release our backdoor triggers and code.

Huawei Lin, Jun Woo Chung, Yingjie Lao et al.

Gradient Boosting Decision Tree (GBDT) is one of the most popular machine learning models in various applications. However, in the traditional settings, all data should be simultaneously accessed in the training procedure: it does not allow to add or delete any data instances after training. In this paper, we propose an efficient online learning framework for GBDT supporting both incremental and decremental learning. To the best of our knowledge, this is the first work that considers an in-place unified incremental and decremental learning on GBDT. To reduce the learning cost, we present a collection of optimizations for our framework, so that it can add or delete a small fraction of data on the fly. We theoretically show the relationship between the hyper-parameters of the proposed optimizations, which enables trading off accuracy and cost on incremental and decremental learning. The backdoor attack results show that our framework can successfully inject and remove backdoor in a well-trained model using incremental and decremental learning, and the empirical results on public datasets confirm the effectiveness and efficiency of our proposed online learning framework and optimizations.

Nicole Meng, Caleb Manicke, David Chen et al.

Adversarial examples represent a serious issue for the application of machine learning models in many sensitive domains. For generating adversarial examples, decision based black-box attacks are one of the most practical techniques as they only require query access to the model. One of the most recently proposed state-of-the-art decision based black-box attacks is Triangle Attack (TA). In this paper, we offer a high-level description of TA and explain potential theoretical limitations. We then propose a new decision based black-box attack, Triangle Attack with Reinforcement Learning (TARL). Our new attack addresses the limits of TA by leveraging reinforcement learning. This creates an attack that can achieve similar, if not better, attack accuracy than TA with half as many queries on state-of-the-art classifiers and defenses across ImageNet and CIFAR-10.

Duy C. Hoang, Hung T. Q. Le, Rui Chu et al.

With the widespread adoption of Large Language Models (LLMs), concerns about potential misuse have emerged. To this end, watermarking has been adapted to LLM, enabling a simple and effective way to detect and monitor generated text. However, while the existing methods can differentiate between watermarked and unwatermarked text with high accuracy, they often face a trade-off between the quality of the generated text and the effectiveness of the watermarking process. In this work, we present a novel type of LLM watermark, Sparse Watermark, which aims to mitigate this trade-off by applying watermarks to a small subset of generated tokens distributed across the text. The key strategy involves anchoring watermarked tokens to words that have specific Part-of-Speech (POS) tags. Our experimental results demonstrate that the proposed watermarking scheme achieves high detectability while generating text that outperforms previous LLM watermarking methods in quality across various tasks

Weihang Tan, Antian Wang, Yingjie Lao et al.

This paper presents a low-latency hardware accelerator for modular polynomial multiplication for lattice-based post-quantum cryptography and homomorphic encryption applications. The proposed novel modular polynomial multiplier exploits the fast finite impulse response (FIR) filter architecture to reduce the computational complexity of the schoolbook modular polynomial multiplication. We also extend this structure to fast $M$-parallel architectures while achieving low-latency, high-speed, and full hardware utilization. We comprehensively evaluate the performance of the proposed architectures under various polynomial settings as well as in the Saber scheme for post-quantum cryptography as a case study. The experimental results show that our proposed modular polynomial multiplier reduces the computation time and area-time product, respectively, compared to the state-of-the-art designs.

Bingyin Zhao, Yingjie Lao

Poisoning attacks on machine learning systems compromise the model performance by deliberately injecting malicious samples in the training dataset to influence the training process. Prior works focus on either availability attacks (i.e., lowering the overall model accuracy) or integrity attacks (i.e., enabling specific instance-based backdoor). In this paper, we advance the adversarial objectives of the availability attacks to a per-class basis, which we refer to as class-oriented poisoning attacks. We demonstrate that the proposed attack is capable of forcing the corrupted model to predict in two specific ways: (i) classify unseen new images to a targeted "supplanter" class, and (ii) misclassify images from a "victim" class while maintaining the classification accuracy on other non-victim classes. To maximize the adversarial effect as well as reduce the computational complexity of poisoned data generation, we propose a gradient-based framework that crafts poisoning images with carefully manipulated feature information for each scenario. Using newly defined metrics at the class level, we demonstrate the effectiveness of the proposed class-oriented poisoning attacks on various models (e.g., LeNet-5, Vgg-9, and ResNet-50) over a wide range of datasets (e.g., MNIST, CIFAR-10, and ImageNet-ILSVRC2012) in an end-to-end training setting.

Joseph Clements, Yuzhe Yang, Ankur Sharma et al.

Recent advances in artificial intelligence and the increasing need for powerful defensive measures in the domain of network security, have led to the adoption of deep learning approaches for use in network intrusion detection systems. These methods have achieved superior performance against conventional network attacks, which enable the deployment of practical security systems to unique and dynamic sectors. Adversarial machine learning, unfortunately, has recently shown that deep learning models are inherently vulnerable to adversarial modifications on their input data. Because of this susceptibility, the deep learning models deployed to power a network defense could in fact be the weakest entry point for compromising a network system. In this paper, we show that by modifying on average as little as 1.38 of the input features, an adversary can generate malicious inputs which effectively fool a deep learning based NIDS. Therefore, when designing such systems, it is crucial to consider the performance from not only the conventional network security perspective but also the adversarial machine learning domain.

Joseph Clements, Yingjie Lao

With the rising popularity of machine learning and the ever increasing demand for computational power, there is a growing need for hardware optimized implementations of neural networks and other machine learning models. As the technology evolves, it is also plausible that machine learning or artificial intelligence will soon become consumer electronic products and military equipment, in the form of well-trained models. Unfortunately, the modern fabless business model of manufacturing hardware, while economic, leads to deficiencies in security through the supply chain. In this paper, we illuminate these security issues by introducing hardware Trojan attacks on neural networks, expanding the current taxonomy of neural network security to incorporate attacks of this nature. To aid in this, we develop a novel framework for inserting malicious hardware Trojans in the implementation of a neural network classifier. We evaluate the capabilities of the adversary in this setting by implementing the attack algorithm on convolutional neural networks while controlling a variety of parameters available to the adversary. Our experimental results show that the proposed algorithm could effectively classify a selected input trigger as a specified class on the MNIST dataset by injecting hardware Trojans into $0.03\%$, on average, of neurons in the 5th hidden layer of arbitrary 7-layer convolutional neural networks, while undetectable under the test data. Finally, we discuss the potential defenses to protect neural networks against hardware Trojan attacks.