Wenyuan Xu

Yancheng Jiang, Yan Jiang, Ruochen Zhou et al.

Virtual Reality (VR) techniques, serving as the bridge between the real and virtual worlds, have boomed and are widely used in manufacturing, remote healthcare, gaming, etc. Specifically, VR systems offer users immersive experiences that include both perceptions and actions. Various studies have demonstrated that attackers can manipulate VR software to influence users' interactions, including perception and actions. However, such attacks typically require strong access and specialized expertise. In this paper, we are the first to present a systematic analysis of physical attacks against VR systems and introduce False Reality, a new attack threat to VR devices without requiring access to or modification of their software. False Reality disturbs VR system services by tampering with sensor measurements, and further spoofing users' perception even inducing harmful actions, e.g., inducing dizziness or causing users to crash into obstacles, by exploiting perceptual and psychological effects. We formalize these threats through an attack pathway framework and validate three representative pathways via physical experiments and user studies on five commercial VR devices. Finally, we further propose a defense prototype to mitigate such threats. Our findings shall provide valuable insights for enhancing the security and resilience of future VR systems.

ByteDance Seed, Jiaze Chen, Tiantian Fan et al. · bytedance

We introduce Seed1.5-Thinking, capable of reasoning through thinking before responding, resulting in improved performance on a wide range of benchmarks. Seed1.5-Thinking achieves 86.7 on AIME 2024, 55.0 on Codeforces and 77.3 on GPQA, demonstrating excellent reasoning abilities in STEM and coding. Beyond reasoning tasks, the method demonstrates notable generalization across diverse domains. For instance, it surpasses DeepSeek R1 by 8% in win rate on non-reasoning tasks, indicating its broader applicability. Compared to other state-of-the-art reasoning models, Seed1.5-Thinking is a Mixture-of-Experts (MoE) model with a relatively small size, featuring 20B activated and 200B total parameters. As part of our effort to assess generalized reasoning, we develop two internal benchmarks, BeyondAIME and Codeforces, both of which will be publicly released to support future research. Model trial link: https://www.volcengine.com/experience/ark.

Xuanwang Zhang, Yunze Song, Yidong Wang et al. · pku

Large Language Models (LLMs) demonstrate human-level capabilities in dialogue, reasoning, and knowledge retention. However, even the most advanced LLMs face challenges such as hallucinations and real-time updating of their knowledge. Current research addresses this bottleneck by equipping LLMs with external knowledge, a technique known as Retrieval Augmented Generation (RAG). However, two key issues constrained the development of RAG. First, there is a growing lack of comprehensive and fair comparisons between novel RAG algorithms. Second, open-source tools such as LlamaIndex and LangChain employ high-level abstractions, which results in a lack of transparency and limits the ability to develop novel algorithms and evaluation metrics. To close this gap, we introduce RAGLAB, a modular and research-oriented open-source library. RAGLAB reproduces 6 existing algorithms and provides a comprehensive ecosystem for investigating RAG algorithms. Leveraging RAGLAB, we conduct a fair comparison of 6 RAG algorithms across 10 benchmarks. With RAGLAB, researchers can efficiently compare the performance of various algorithms and develop novel algorithms.

Hanjun Luo, Haoyu Huang, Ziye Deng et al.

Text-to-Image (T2I) generative models are becoming increasingly crucial due to their ability to generate high-quality images, but also raise concerns about social biases, particularly in human image generation. Sociological research has established systematic classifications of bias. Yet, existing studies on bias in T2I models largely conflate different types of bias, impeding methodological progress. In this paper, we introduce BIGbench, a unified benchmark for Biases of Image Generation, featuring a carefully designed dataset. Unlike existing benchmarks, BIGbench classifies and evaluates biases across four dimensions to enable a more granular evaluation and deeper analysis. Furthermore, BIGbench applies advanced multi-modal large language models to achieve fully automated and highly accurate evaluations. We apply BIGbench to evaluate eight representative T2I models and three debiasing methods. Our human evaluation results by trained evaluators from different races underscore BIGbench's effectiveness in aligning images and identifying various biases. Moreover, our study also reveals new research directions about biases with insightful analysis of our results. Our work is openly accessible at https://github.com/BIGbench2024/BIGbench2024/.

Chen Yan, Zhijian Xu, Zhanyuan Yin et al.

Traffic light recognition is essential for fully autonomous driving in urban areas. In this paper, we investigate the feasibility of fooling traffic light recognition mechanisms by shedding laser interference on the camera. By exploiting the rolling shutter of CMOS sensors, we manage to inject a color stripe overlapped on the traffic light in the image, which can cause a red light to be recognized as a green light or vice versa. To increase the success rate, we design an optimization method to search for effective laser parameters based on empirical models of laser interference. Our evaluation in emulated and real-world setups on 2 state-of-the-art recognition systems and 5 cameras reports a maximum success rate of 30% and 86.25% for Red-to-Green and Green-to-Red attacks. We observe that the attack is effective in continuous frames from more than 40 meters away against a moving vehicle, which may cause end-to-end impacts on self-driving such as running a red light or emergency stop. To mitigate the threat, we propose redesigning the rolling shutter mechanism.

Xinfeng Li, Kai Li, Yifan Zheng et al.

Text-to-Speech (TTS) and Voice Conversion (VC) models have exhibited remarkable performance in generating realistic and natural audio. However, their dark side, audio deepfake poses a significant threat to both society and individuals. Existing countermeasures largely focus on determining the genuineness of speech based on complete original audio recordings, which however often contain private content. This oversight may refrain deepfake detection from many applications, particularly in scenarios involving sensitive information like business secrets. In this paper, we propose SafeEar, a novel framework that aims to detect deepfake audios without relying on accessing the speech content within. Our key idea is to devise a neural audio codec into a novel decoupling model that well separates the semantic and acoustic information from audio samples, and only use the acoustic information (e.g., prosody and timbre) for deepfake detection. In this way, no semantic content will be exposed to the detector. To overcome the challenge of identifying diverse deepfake audio without semantic clues, we enhance our deepfake detector with real-world codec augmentation. Extensive experiments conducted on four benchmark datasets demonstrate SafeEar's effectiveness in detecting various deepfake techniques with an equal error rate (EER) down to 2.02%. Simultaneously, it shields five-language speech content from being deciphered by both machine and human auditory analysis, demonstrated by word error rates (WERs) all above 93.93% and our user study. Furthermore, our benchmark constructed for anti-deepfake and anti-content recovery evaluation helps provide a basis for future research in the realms of audio privacy preservation and deepfake detection.

Zizhi Jin, Qinhong Jiang, Xuancun Lu et al.

LiDAR (Light Detection and Ranging) is a pivotal sensor for autonomous driving, offering precise 3D spatial information. Previous signal attacks against LiDAR systems mainly exploit laser signals. In this paper, we investigate the possibility of cross-modality signal injection attacks, i.e., injecting intentional electromagnetic interference (IEMI) to manipulate LiDAR output. Our insight is that the internal modules of a LiDAR, i.e., the laser receiving circuit, the monitoring sensors, and the beam-steering modules, even with strict electromagnetic compatibility (EMC) testing, can still couple with the IEMI attack signals and result in the malfunction of LiDAR systems. Based on the above attack surfaces, we propose the PhantomLiDAR attack, which manipulates LiDAR output in terms of Points Interference, Points Injection, Points Removal, and even LiDAR Power-Off. We evaluate and demonstrate the effectiveness of PhantomLiDAR with both simulated and real-world experiments on five COTS LiDAR systems. We also conduct feasibility experiments in real-world moving scenarios. We provide potential defense measures that can be implemented at both the sensor level and the vehicle system level to mitigate the risks associated with IEMI attacks. Video demonstrations can be viewed at https://sites.google.com/view/phantomlidar.

Yan Long, Chen Yan, Shilin Xiao et al.

Using mathematical modeling and human subjects experiments, this research explores the extent to which emerging webcams might leak recognizable textual and graphical information gleaming from eyeglass reflections captured by webcams. The primary goal of our work is to measure, compute, and predict the factors, limits, and thresholds of recognizability as webcam technology evolves in the future. Our work explores and characterizes the viable threat models based on optical attacks using multi-frame super resolution techniques on sequences of video frames. Our models and experimental results in a controlled lab setting show it is possible to reconstruct and recognize with over 75% accuracy on-screen texts that have heights as small as 10 mm with a 720p webcam. We further apply this threat model to web textual contents with varying attacker capabilities to find thresholds at which text becomes recognizable. Our user study with 20 participants suggests present-day 720p webcams are sufficient for adversaries to reconstruct textual content on big-font websites. Our models further show that the evolution towards 4K cameras will tip the threshold of text leakage to reconstruction of most header texts on popular websites. Besides textual targets, a case study on recognizing a closed-world dataset of Alexa top 100 websites with 720p webcams shows a maximum recognition accuracy of 94% with 10 participants even without using machine-learning models. Our research proposes near-term mitigations including a software prototype that users can use to blur the eyeglass areas of their video streams. For possible long-term defenses, we advocate an individual reflection testing procedure to assess threats under various settings, and justify the importance of following the principle of least privilege for privacy-sensitive scenarios.

Jiangyi Deng, Yanjiao Chen, Yinan Zhong et al.

Voice conversion (VC) techniques can be abused by malicious parties to transform their audios to sound like a target speaker, making it hard for a human being or a speaker verification/identification system to trace the source speaker. In this paper, we make the first attempt to restore the source voiceprint from audios synthesized by voice conversion methods with high credit. However, unveiling the features of the source speaker from a converted audio is challenging since the voice conversion operation intends to disentangle the original features and infuse the features of the target speaker. To fulfill our goal, we develop Revelio, a representation learning model, which learns to effectively extract the voiceprint of the source speaker from converted audio samples. We equip Revelio with a carefully-designed differential rectification algorithm to eliminate the influence of the target speaker by removing the representation component that is parallel to the voiceprint of the target speaker. We have conducted extensive experiments to evaluate the capability of Revelio in restoring voiceprint from audios converted by VQVC, VQVC+, AGAIN, and BNE. The experiments verify that Revelio is able to rebuild voiceprints that can be traced to the source speaker by speaker verification and identification systems. Revelio also exhibits robust performance under inter-gender conversion, unseen languages, and telephony networks.

Youqian Zhang, Chunxi Yang, Eugene Y. Fu et al.

Object detection can localize and identify objects in images, and it is extensively employed in critical multimedia applications such as security surveillance and autonomous driving. Despite the success of existing object detection models, they are often evaluated in ideal scenarios where captured images guarantee the accurate and complete representation of the detecting scenes. However, images captured by image sensors may be affected by different factors in real applications, including cyber-physical attacks. In particular, attackers can exploit hardware properties within the systems to inject electromagnetic interference so as to manipulate the images. Such attacks can cause noisy or incomplete information about the captured scene, leading to incorrect detection results, potentially granting attackers malicious control over critical functions of the systems. This paper presents a research work that comprehensively quantifies and analyzes the impacts of such attacks on state-of-the-art object detection models in practice. It also sheds light on the underlying reasons for the incorrect detection outcomes.

Jiangyi Deng, Xinfeng Li, Yanjiao Chen et al.

Malicious shell commands are linchpins to many cyber-attacks, but may not be easy to understand by security analysts due to complicated and often disguised code structures. Advances in large language models (LLMs) have unlocked the possibility of generating understandable explanations for shell commands. However, existing general-purpose LLMs suffer from a lack of expert knowledge and a tendency to hallucinate in the task of shell command explanation. In this paper, we present Raconteur, a knowledgeable, expressive and portable shell command explainer powered by LLM. Raconteur is infused with professional knowledge to provide comprehensive explanations on shell commands, including not only what the command does (i.e., behavior) but also why the command does it (i.e., purpose). To shed light on the high-level intent of the command, we also translate the natural-language-based explanation into standard technique & tactic defined by MITRE ATT&CK, the worldwide knowledge base of cybersecurity. To enable Raconteur to explain unseen private commands, we further develop a documentation retriever to obtain relevant information from complementary documentations to assist the explanation process. We have created a large-scale dataset for training and conducted extensive experiments to evaluate the capability of Raconteur in shell command explanation. The experiments verify that Raconteur is able to provide high-quality explanations and in-depth insight of the intent of the command.

Jialin Wu, Jiangyi Deng, Shengyuan Pang et al.

Given the societal impact of unsafe content generated by large language models (LLMs), ensuring that LLM services comply with safety standards is a crucial concern for LLM service providers. Common content moderation methods are limited by an effectiveness-and-efficiency dilemma, where simple models are fragile while sophisticated models consume excessive computational resources. In this paper, we reveal for the first time that effective and efficient content moderation can be achieved by extracting conceptual features from chat-oriented LLMs, despite their initial fine-tuning for conversation rather than content moderation. We propose a practical and unified content moderation framework for LLM services, named Legilimens, which features both effectiveness and efficiency. Our red-team model-based data augmentation enhances the robustness of Legilimens against state-of-the-art jailbreaking. Additionally, we develop a framework to theoretically analyze the cost-effectiveness of Legilimens compared to other methods. We have conducted extensive experiments on five host LLMs, seventeen datasets, and nine jailbreaking methods to verify the effectiveness, efficiency, and robustness of Legilimens against normal and adaptive adversaries. A comparison of Legilimens with both commercial and academic baselines demonstrates the superior performance of Legilimens. Furthermore, we confirm that Legilimens can be applied to few-shot scenarios and extended to multi-label classification tasks.

Xuancun Lu, Jiaxiang Chen, Shilin Xiao et al.

Vision-Language-Action (VLA) models revolutionize robotic systems by enabling end-to-end perception-to-action pipelines that integrate multiple sensory modalities, such as visual signals processed by cameras and auditory signals captured by microphones. This multi-modality integration allows VLA models to interpret complex, real-world environments using diverse sensor data streams. Given the fact that VLA-based systems heavily rely on the sensory input, the security of VLA models against physical-world sensor attacks remains critically underexplored. To address this gap, we present the first systematic study of physical sensor attacks against VLAs, quantifying the influence of sensor attacks and investigating the defenses for VLA models. We introduce a novel ``Real-Sim-Real'' framework that automatically simulates physics-based sensor attack vectors, including six attacks targeting cameras and two targeting microphones, and validates them on real robotic systems. Through large-scale evaluations across various VLA architectures and tasks under varying attack parameters, we demonstrate significant vulnerabilities, with susceptibility patterns that reveal critical dependencies on task types and model designs. We further develop an adversarial-training-based defense that enhances VLA robustness against out-of-distribution physical perturbations caused by sensor attacks while preserving model performance. Our findings expose an urgent need for standardized robustness benchmarks and mitigation strategies to secure VLA deployments in safety-critical environments.

Siyi Chen, Tianhan Luo, Shijian Wu et al.

Open-source libraries are widely used in modern software development, introducing significant security vulnerabilities. While static analysis tools can identify potential vulnerabilities at scale, they often generate overwhelming reports with high false positive rates. Automated Exploit Generation (AEG) emerges as a promising solution to confirm vulnerability authenticity by generating an exploit. However, traditional AEG approaches based on fuzzing or symbolic execution face path coverage and constraint-solving problems. Although LLMs show great potential for AEG, how to effectively leverage them to comprehend vulnerabilities and generate corresponding exploits is still an open question. To address these challenges, we propose Vulnsage, a multi-agent framework for AEG. Vulnsage simulates human security researchers' workflows by decomposing the complex AEG process into multiple specialized sub-agents: Code Analyzer Agent, Code Generation Agent, Validation Agent, and a set of Reflection Agents, orchestrated by a central supervisor through iterative cycles. Given a target program, the Code Analyzer Agent performs static analysis to identify potential vulnerabilities and collects relevant information for each one. The Code Generation Agent then utilizes an LLM to generate candidate exploits. The Validation Agent and Reflection Agents form a feedback-driven self-refinement loop that uses execution traces and runtime error analysis to either improve the exploit iteratively or reason about the false positive alert. Experimental evaluation demonstrates that Vulnsage succeeds in generating 34.64\% more exploits than state-of-the-art tools such as \explodejs. Furthermore, Vulnsage has successfully discovered and verified 146 zero-day vulnerabilities in real-world scenarios, demonstrating its practical effectiveness for assisting security assessment in software supply chains.

Yujie Gu, Richeng Jin, Xiaoyu Ji et al.

Large Language Models (LLMs) have achieved remarkable performance and received significant research interest. The enormous computational demands, however, hinder the local deployment on devices with limited resources. The current prevalent LLM inference paradigms require users to send queries to the service providers for processing, which raises critical privacy concerns. Existing approaches propose to allow the users to obfuscate the token embeddings before transmission and utilize local models for denoising. Nonetheless, transmitting the token embeddings and deploying local models may result in excessive communication and computation overhead, preventing practical implementation. In this work, we propose \textbf{DEL}, a framework for \textbf{D}ifferentially private and communication \textbf{E}fficient \textbf{L}LM split inference. More specifically, an embedding projection module and a differentially private stochastic quantization mechanism are proposed to reduce the communication overhead in a privacy-preserving manner. To eliminate the need for local models, we adapt soft prompt at the server side to compensate for the utility degradation caused by privacy. To the best of our knowledge, this is the first work that utilizes soft prompt to improve the trade-off between privacy and utility in LLM inference, and extensive experiments on text generation and natural language understanding benchmarks demonstrate the effectiveness of the proposed method.

Hanjun Luo, Chiming Ni, Jiaheng Wen et al.

LLM-powered coding agents are reshaping the development paradigm. However, existing evaluation systems, neither traditional tests for humans nor benchmarks for LLMs, fail to capture this shift. They remain focused on well-defined algorithmic problems, which excludes problems where success depends on human-AI collaboration. Such collaborative problems not only require human reasoning to interpret complex contexts and guide solution strategies, but also demand AI efficiency for implementation. To bridge this gap, we introduce HAI-Eval, a unified benchmark designed to measure the synergy of human-AI partnership in coding. HAI-Eval's core innovation is its "Collaboration-Necessary" problem templates, which are intractable for both standalone LLMs and unaided humans, but solvable through effective collaboration. Specifically, HAI-Eval uses 45 templates to dynamically create tasks. It also provides a standardized IDE for human participants and a reproducible toolkit with 450 task instances for LLMs, ensuring an ecologically valid evaluation. We conduct a within-subject study with 45 participants and benchmark their performance against 5 state-of-the-art LLMs under 4 different levels of human intervention. Results show that standalone LLMs and unaided participants achieve poor pass rates (0.67% and 18.89%), human-AI collaboration significantly improves performance to 31.11%. Our analysis reveals an emerging co-reasoning partnership. This finding challenges the traditional human-tool hierarchy by showing that strategic breakthroughs can originate from either humans or AI. HAI-Eval establishes not only a challenging benchmark for next-generation coding agents but also a grounded, scalable framework for assessing core developer competencies in the AI era. Our benchmark and interactive demo will be openly accessible.

Kai Li, Can Shen, Yile Liu et al.

Audio Large Language Models (ALLMs) have gained widespread adoption, yet their trustworthiness remains underexplored. Existing evaluation frameworks, designed primarily for text, fail to address unique vulnerabilities introduced by audio's acoustic properties. We identify significant trustworthiness risks in ALLMs arising from non-semantic acoustic cues, including timbre, accent, and background noise, which can manipulate model behavior. We propose AudioTrust, a comprehensive framework for systematic evaluation of ALLM trustworthiness across audio-specific risks. AudioTrust encompasses six key dimensions: fairness, hallucination, safety, privacy, robustness, and authentication. The framework implements 26 distinct sub-tasks using a curated dataset of over 4,420 audio samples from real-world scenarios, including daily conversations, emergency calls, and voice assistant interactions. We conduct comprehensive evaluations across 18 experimental configurations using human-validated automated pipelines. Our evaluation of 14 state-of-the-art open-source and closed-source ALLMs reveals significant limitations when confronted with diverse high-risk audio scenarios, providing insights for secure deployment of audio models. Code and data are available at https://github.com/JusperLee/AudioTrust.

Zhengxian Huang, Wenjun Zhu, Haoxuan Qiu et al.

By integrating Chain-of-Thought(CoT) reasoning, Vision-Language-Action (VLA) models have demonstrated strong capabilities in robotic manipulation, particularly by improving generalization and interpretability. However, the security of CoT-based reasoning mechanisms remains largely unexplored. In this paper, we show that CoT reasoning introduces a novel attack vector for targeted control hijacking--for example, causing a robot to mistakenly deliver a knife to a person instead of an apple--without modifying the user's instruction. We first provide empirical evidence that CoT strongly governs action generation, even when it is semantically misaligned with the input instructions. Building on this observation, we propose TRAP, the first targeted adversarial attack framework for CoT-reasoning VLA models. TRAP uses an adversarial patch (e.g., a coaster placed on the table) to corrupt intermediate CoT reasoning and hijack the VLA's output. By optimizing the CoT adversarial loss, TRAP induces specific and adversary-defined behaviors. Extensive evaluations across 3 mainstream VLA architectures and 3 CoT reasoning paradigms validate the effectiveness of TRAP. Notably, we implemented the patch by printing it on paper in a real-world setting. Our findings highlight the urgent need to secure CoT reasoning in VLA systems.

Yu Yue, Yufeng Yuan, Qiying Yu et al.

We present VAPO, Value-based Augmented Proximal Policy Optimization framework for reasoning models., a novel framework tailored for reasoning models within the value-based paradigm. Benchmarked the AIME 2024 dataset, VAPO, built on the Qwen 32B pre-trained model, attains a state-of-the-art score of $\mathbf{60.4}$. In direct comparison under identical experimental settings, VAPO outperforms the previously reported results of DeepSeek-R1-Zero-Qwen-32B and DAPO by more than 10 points. The training process of VAPO stands out for its stability and efficiency. It reaches state-of-the-art performance within a mere 5,000 steps. Moreover, across multiple independent runs, no training crashes occur, underscoring its reliability. This research delves into long chain-of-thought (long-CoT) reasoning using a value-based reinforcement learning framework. We pinpoint three key challenges that plague value-based methods: value model bias, the presence of heterogeneous sequence lengths, and the sparsity of reward signals. Through systematic design, VAPO offers an integrated solution that effectively alleviates these challenges, enabling enhanced performance in long-CoT reasoning tasks.

Kun Wang, Guibin Zhang, Zhenhong Zhou et al. · mit

The remarkable success of Large Language Models (LLMs) has illuminated a promising pathway toward achieving Artificial General Intelligence for both academic and industrial communities, owing to their unprecedented performance across various applications. As LLMs continue to gain prominence in both research and commercial domains, their security and safety implications have become a growing concern, not only for researchers and corporations but also for every nation. Currently, existing surveys on LLM safety primarily focus on specific stages of the LLM lifecycle, e.g., deployment phase or fine-tuning phase, lacking a comprehensive understanding of the entire "lifechain" of LLMs. To address this gap, this paper introduces, for the first time, the concept of "full-stack" safety to systematically consider safety issues throughout the entire process of LLM training, deployment, and eventual commercialization. Compared to the off-the-shelf LLM safety surveys, our work demonstrates several distinctive advantages: (I) Comprehensive Perspective. We define the complete LLM lifecycle as encompassing data preparation, pre-training, post-training, deployment and final commercialization. To our knowledge, this represents the first safety survey to encompass the entire lifecycle of LLMs. (II) Extensive Literature Support. Our research is grounded in an exhaustive review of over 800+ papers, ensuring comprehensive coverage and systematic organization of security issues within a more holistic understanding. (III) Unique Insights. Through systematic literature analysis, we have developed reliable roadmaps and perspectives for each chapter. Our work identifies promising research directions, including safety in data generation, alignment techniques, model editing, and LLM-based agent systems. These insights provide valuable guidance for researchers pursuing future work in this field.

Xinfeng Li, Yuchen Yang, Jiangyi Deng et al.

Text-to-image (T2I) models, such as Stable Diffusion, have exhibited remarkable performance in generating high-quality images from text descriptions in recent years. However, text-to-image models may be tricked into generating not-safe-for-work (NSFW) content, particularly in sexually explicit scenarios. Existing countermeasures mostly focus on filtering inappropriate inputs and outputs, or suppressing improper text embeddings, which can block sexually explicit content (e.g., naked) but may still be vulnerable to adversarial prompts -- inputs that appear innocent but are ill-intended. In this paper, we present SafeGen, a framework to mitigate sexual content generation by text-to-image models in a text-agnostic manner. The key idea is to eliminate explicit visual representations from the model regardless of the text input. In this way, the text-to-image model is resistant to adversarial prompts since such unsafe visual representations are obstructed from within. Extensive experiments conducted on four datasets and large-scale user studies demonstrate SafeGen's effectiveness in mitigating sexually explicit content generation while preserving the high-fidelity of benign images. SafeGen outperforms eight state-of-the-art baseline methods and achieves 99.4% sexual content removal performance. Furthermore, our constructed benchmark of adversarial prompts provides a basis for future development and evaluation of anti-NSFW-generation methods.

Wenjun Zhu, Xiaoyu Ji, Yushi Cheng et al.

Autonomous vehicles increasingly utilize the vision-based perception module to acquire information about driving environments and detect obstacles. Correct detection and classification are important to ensure safe driving decisions. Existing works have demonstrated the feasibility of fooling the perception models such as object detectors and image classifiers with printed adversarial patches. However, most of them are indiscriminately offensive to every passing autonomous vehicle. In this paper, we propose TPatch, a physical adversarial patch triggered by acoustic signals. Unlike other adversarial patches, TPatch remains benign under normal circumstances but can be triggered to launch a hiding, creating or altering attack by a designed distortion introduced by signal injection attacks towards cameras. To avoid the suspicion of human drivers and make the attack practical and robust in the real world, we propose a content-based camouflage method and an attack robustness enhancement method to strengthen it. Evaluations with three object detectors, YOLO V3/V5 and Faster R-CNN, and eight image classifiers demonstrate the effectiveness of TPatch in both the simulation and the real world. We also discuss possible defenses at the sensor, algorithm, and system levels.

Wenyuan Xu, Xiaochen Zuo, Chao Xin et al.

Reinforcement Learning from Human Feedback (RLHF) has emerged as a important paradigm for aligning large language models (LLMs) with human preferences during post-training. This framework typically involves two stages: first, training a reward model on human preference data, followed by optimizing the language model using reinforcement learning algorithms. However, current RLHF approaches may constrained by two limitations. First, existing RLHF frameworks often rely on Bradley-Terry models to assign scalar rewards based on pairwise comparisons of individual responses. However, this approach imposes significant challenges on reward model (RM), as the inherent variability in prompt-response pairs across different contexts demands robust calibration capabilities from the RM. Second, reward models are typically initialized from generative foundation models, such as pre-trained or supervised fine-tuned models, despite the fact that reward models perform discriminative tasks, creating a mismatch. This paper introduces Pairwise-RL, a RLHF framework that addresses these challenges through a combination of generative reward modeling and a pairwise proximal policy optimization (PPO) algorithm. Pairwise-RL unifies reward model training and its application during reinforcement learning within a consistent pairwise paradigm, leveraging generative modeling techniques to enhance reward model performance and score calibration. Experimental evaluations demonstrate that Pairwise-RL outperforms traditional RLHF frameworks across both internal evaluation datasets and standard public benchmarks, underscoring its effectiveness in improving alignment and model behavior.

Jiangyi Deng, Shengyuan Pang, Yanjiao Chen et al.

Instead of building deep learning models from scratch, developers are more and more relying on adapting pre-trained models to their customized tasks. However, powerful pre-trained models may be misused for unethical or illegal tasks, e.g., privacy inference and unsafe content generation. In this paper, we introduce a pioneering learning paradigm, non-fine-tunable learning, which prevents the pre-trained model from being fine-tuned to indecent tasks while preserving its performance on the original task. To fulfill this goal, we propose SOPHON, a protection framework that reinforces a given pre-trained model to be resistant to being fine-tuned in pre-defined restricted domains. Nonetheless, this is challenging due to a diversity of complicated fine-tuning strategies that may be adopted by adversaries. Inspired by model-agnostic meta-learning, we overcome this difficulty by designing sophisticated fine-tuning simulation and fine-tuning evaluation algorithms. In addition, we carefully design the optimization process to entrap the pre-trained model within a hard-to-escape local optimum regarding restricted domains. We have conducted extensive experiments on two deep learning modes (classification and generation), seven restricted domains, and six model architectures to verify the effectiveness of SOPHON. Experiment results verify that fine-tuning SOPHON-protected models incurs an overhead comparable to or even greater than training from scratch. Furthermore, we confirm the robustness of SOPHON to three fine-tuning methods, five optimizers, various learning rates and batch sizes. SOPHON may help boost further investigations into safe and responsible AI.

Yuhua Jiang, Yuwen Xiong, Yufeng Yuan et al.

Large Language Models (LLMs) have demonstrated impressive capabilities in complex reasoning tasks, yet they still struggle to reliably verify the correctness of their own outputs. Existing solutions to this verification challenge often depend on separate verifier models or require multi-stage self-correction training pipelines, which limit scalability. In this paper, we propose Policy as Generative Verifier (PAG), a simple and effective framework that empowers LLMs to self-correct by alternating between policy and verifier roles within a unified multi-turn reinforcement learning (RL) paradigm. Distinct from prior approaches that always generate a second attempt regardless of model confidence, PAG introduces a selective revision mechanism: the model revises its answer only when its own generative verification step detects an error. This verify-then-revise workflow not only alleviates model collapse but also jointly enhances both reasoning and verification abilities. Extensive experiments across diverse reasoning benchmarks highlight PAG's dual advancements: as a policy, it enhances direct generation and self-correction accuracy; as a verifier, its self-verification outperforms self-consistency.

Xuancun Lu, Zhengxian Huang, Xinfeng Li et al.

The integration of LLMs into robots has witnessed significant growth, where LLMs can convert instructions into executable robot policies. However, the inherent vulnerability of LLMs to jailbreak attacks brings critical security risks from the digital domain to the physical world. An attacked LLM-based robot could execute harmful policies and cause physical harm. In this paper, we investigate the feasibility and rationale of jailbreak attacks against LLM-based robots and answer three research questions: (1) How applicable are existing LLM jailbreak attacks against LLM-based robots? (2) What unique challenges arise if they are not directly applicable? (3) How to defend against such jailbreak attacks? To this end, we first construct a "human-object-environment" robot risks-oriented Harmful-RLbench and then conduct a measurement study on LLM-based robot systems. Our findings conclude that traditional LLM jailbreak attacks are inapplicable in robot scenarios, and we identify two unique challenges: determining policy-executable optimization directions and accurately evaluating robot-jailbroken policies. To enable a more thorough security analysis, we introduce POEX (POlicy EXecutable) jailbreak, a red-teaming framework that induces harmful yet executable policy to jailbreak LLM-based robots. POEX incorporates hidden layer gradient optimization to guarantee jailbreak success and policy execution as well as a multi-agent evaluator to accurately assess the practical executability of policies. Experiments conducted on the real-world robotic systems and in simulation demonstrate the efficacy of POEX, highlighting critical security vulnerabilities and its transferability across LLMs. Finally, we propose prompt-based and model-based defenses to mitigate attacks. Our findings underscore the urgent need for security measures to ensure the safe deployment of LLM-based robots in critical applications.

Wenjun Zhu, Yuan Sun, Jiani Liu et al.

The proliferation of images captured from millions of cameras and the advancement of facial recognition (FR) technology have made the abuse of FR a severe privacy threat. Existing works typically rely on obfuscation, synthesis, or adversarial examples to modify faces in images to achieve anti-facial recognition (AFR). However, the unmodified images captured by camera modules that contain sensitive personally identifiable information (PII) could still be leaked. In this paper, we propose a novel approach, CamPro, to capture inborn AFR images. CamPro enables well-packed commodity camera modules to produce images that contain little PII and yet still contain enough information to support other non-sensitive vision applications, such as person detection. Specifically, CamPro tunes the configuration setup inside the camera image signal processor (ISP), i.e., color correction matrix and gamma correction, to achieve AFR, and designs an image enhancer to keep the image quality for possible human viewers. We implemented and validated CamPro on a proof-of-concept camera, and our experiments demonstrate its effectiveness on ten state-of-the-art black-box FR models. The results show that CamPro images can significantly reduce face identification accuracy to 0.3\% while having little impact on the targeted non-sensitive vision application. Furthermore, we find that CamPro is resilient to adaptive attackers who have re-trained their FR models using images generated by CamPro, even with full knowledge of privacy-preserving ISP parameters.

Jialin Wu, Kaikai Pan, Yanjiao Chen et al.

Transformer models have excelled in natural language tasks, prompting the vision community to explore their implementation in computer vision problems. However, these models are still influenced by adversarial examples. In this paper, we investigate the attack capabilities of six common adversarial attacks on three pretrained ViT models to reveal the vulnerability of ViT models. To understand and analyse the bias in neural network decisions when the input is adversarial, we use two visualisation techniques that are attention rollout and grad attention rollout. To prevent ViT models from adversarial attack, we propose Protego, a detection framework that leverages the transformer intrinsic capabilities to detection adversarial examples of ViT models. Nonetheless, this is challenging due to a diversity of attack strategies that may be adopted by adversaries. Inspired by the attention mechanism, we know that the token of prediction contains all the information from the input sample. Additionally, the attention region for adversarial examples differs from that of normal examples. Given these points, we can train a detector that achieves superior performance than existing detection methods to identify adversarial examples. Our experiments have demonstrated the high effectiveness of our detection method. For these six adversarial attack methods, our detector's AUC scores all exceed 0.95. Protego may advance investigations in metaverse security.

Xinfeng Li, Shengyuan Pang, Jialin Wu et al.

Text-to-image (T2I) models, though exhibiting remarkable creativity in image generation, can be exploited to produce unsafe images. Existing safety measures, e.g., content moderation or model alignment, fail in the presence of white-box adversaries who know and can adjust model parameters, e.g., by fine-tuning. This paper presents a novel defensive framework, named Patronus, which equips T2I models with holistic protection to defend against white-box adversaries. Specifically, we design an internal moderator that decodes unsafe input features into zero vectors while ensuring the decoding performance of benign input features. Furthermore, we strengthen the model alignment with a carefully designed non-fine-tunable learning mechanism, ensuring the T2I model will not be compromised by malicious fine-tuning. We conduct extensive experiments to validate the intactness of the performance on safe content generation and the effectiveness of rejecting unsafe content generation. Results also confirm the resilience of Patronus against various fine-tuning attacks by white-box adversaries.

Jiangyi Deng, Fei Teng, Yanjiao Chen et al.

Voice data generated on instant messaging or social media applications contains unique user voiceprints that may be abused by malicious adversaries for identity inference or identity theft. Existing voice anonymization techniques, e.g., signal processing and voice conversion/synthesis, suffer from degradation of perceptual quality. In this paper, we develop a voice anonymization system, named V-Cloak, which attains real-time voice anonymization while preserving the intelligibility, naturalness and timbre of the audio. Our designed anonymizer features a one-shot generative model that modulates the features of the original audio at different frequency levels. We train the anonymizer with a carefully-designed loss function. Apart from the anonymity loss, we further incorporate the intelligibility loss and the psychoacoustics-based naturalness loss. The anonymizer can realize untargeted and targeted anonymization to achieve the anonymity goals of unidentifiability and unlinkability. We have conducted extensive experiments on four datasets, i.e., LibriSpeech (English), AISHELL (Chinese), CommonVoice (French) and CommonVoice (Italian), five Automatic Speaker Verification (ASV) systems (including two DNN-based, two statistical and one commercial ASV), and eleven Automatic Speech Recognition (ASR) systems (for different languages). Experiment results confirm that V-Cloak outperforms five baselines in terms of anonymity performance. We also demonstrate that V-Cloak trained only on the VoxCeleb1 dataset against ECAPA-TDNN ASV and DeepSpeech2 ASR has transferable anonymity against other ASVs and cross-language intelligibility for other ASRs. Furthermore, we verify the robustness of V-Cloak against various de-noising techniques and adaptive attacks. Hopefully, V-Cloak may provide a cloak for us in a prism world.

Yanjiao Chen, Yijie Bai, Richard Mitev et al.

In the area of Internet of Things (IoT) voice assistants have become an important interface to operate smart speakers, smartphones, and even automobiles. To save power and protect user privacy, voice assistants send commands to the cloud only if a small set of pre-registered wake-up words are detected. However, voice assistants are shown to be vulnerable to the FakeWake phenomena, whereby they are inadvertently triggered by innocent-sounding fuzzy words. In this paper, we present a systematic investigation of the FakeWake phenomena from three aspects. To start with, we design the first fuzzy word generator to automatically and efficiently produce fuzzy words instead of searching through a swarm of audio materials. We manage to generate 965 fuzzy words covering 8 most popular English and Chinese smart speakers. To explain the causes underlying the FakeWake phenomena, we construct an interpretable tree-based decision model, which reveals phonetic features that contribute to false acceptance of fuzzy words by wake-up word detectors. Finally, we propose remedies to mitigate the effect of FakeWake. The results show that the strengthened models are not only resilient to fuzzy words but also achieve better overall performance on original training datasets.

Weizhao Jin, Xiaoyu Ji, Ruiwen He et al.

With the rapid growth of online services, the number of online accounts proliferates. The security of a single user account no longer depends merely on its own service provider but also the accounts on other service platforms(We refer to this online account environment as Online Account Ecosystem). In this paper, we first uncover the vulnerability of Online Account Ecosystem, which stems from the defective multi-factor authentication (MFA), specifically the ones with SMS-based verification, and dependencies among accounts on different platforms. We propose Chain Reaction Attack that exploits the weakest point in Online Account Ecosystem and can ultimately compromise the most secure platform. Furthermore, we design and implement ActFort, a systematic approach to detect the vulnerability of Online Account Ecosystem by analyzing the authentication credential factors and sensitive personal information as well as evaluating the dependency relationships among online accounts. We evaluate our system on hundreds of representative online services listed in Alexa in diversified fields. Based on the analysis from ActFort, we provide several pragmatic insights into the current Online Account Ecosystem and propose several feasible countermeasures including the online account exposed information protection mechanism and the built-in authentication to fortify the security of Online Account Ecosystem.

Xiu-Shen Wei, Yu-Yan Xu, Yazhou Yao et al.

WebFG 2020 is an international challenge hosted by Nanjing University of Science and Technology, University of Edinburgh, Nanjing University, The University of Adelaide, Waseda University, etc. This challenge mainly pays attention to the webly-supervised fine-grained recognition problem. In the literature, existing deep learning methods highly rely on large-scale and high-quality labeled training data, which poses a limitation to their practicability and scalability in real world applications. In particular, for fine-grained recognition, a visual task that requires professional knowledge for labeling, the cost of acquiring labeled training data is quite high. It causes extreme difficulties to obtain a large amount of high-quality training data. Therefore, utilizing free web data to train fine-grained recognition models has attracted increasing attentions from researchers in the fine-grained community. This challenge expects participants to develop webly-supervised fine-grained recognition methods, which leverages web images in training fine-grained recognition models to ease the extreme dependence of deep learning methods on large-scale manually labeled datasets and to enhance their practicability and scalability. In this technical report, we have pulled together the top WebFG 2020 solutions of total 54 competing teams, and discuss what methods worked best across the set of winning teams, and what surprisingly did not help.

Mengran Liu, Weiwei Fang, Xiaodong Ma et al.

Deep Neural Networks (DNNs) have achieved remarkable success in many computer vision tasks recently, but the huge number of parameters and the high computation overhead hinder their deployments on resource-constrained edge devices. It is worth noting that channel pruning is an effective approach for compressing DNN models. A critical challenge is to determine which channels are to be removed, so that the model accuracy will not be negatively affected. In this paper, we first propose Spatial and Channel Attention (SCA), a new attention module combining both spatial and channel attention that respectively focuses on "where" and "what" are the most informative parts. Guided by the scale values generated by SCA for measuring channel importance, we further propose a new channel pruning approach called Channel Pruning guided by Spatial and Channel Attention (CPSCA). Experimental results indicate that SCA achieves the best inference accuracy, while incurring negligibly extra resource consumption, compared to other state-of-the-art attention modules. Our evaluation on two benchmark datasets shows that, with the guidance of SCA, our CPSCA approach achieves higher inference accuracy than other state-of-the-art pruning methods under the same pruning ratios.

Zhiyun Wang, Jiayu Zhang, Xiaoyu Ji et al.

With the increasing popularity of the Internet of Things(IoT) devices, the demand for fast and convenient battery charging services grows rapidly. Wireless charging is a promising technology for such a purpose and its usage has become ubiquitous. However, the close distance between the charger and the device being charged not only makes proximity-based and near field communication attacks possible, but also introduces a new type of vulnerabilities. In this paper, we propose to create fingerprints for wireless chargers based on the intrinsic non-linear distortion effects of the underlying charging circuit. Using such fingerprints, we design the WirelessID system to detect potential short-range malicious wireless charging attacks. WirelessID collects signals in the standby state of the charging process and sends them to a trusted server, which can extract the fingerprint and then identify the charger.

Wenyuan Xu, Xiaopeng Li, Jing Tian et al.

With the emerging of touch-less human-computer interaction techniques and gadgets, mid-air hand gestures have been widely used for authentication. Much literature examined either the usability or security of a handful of gestures. This paper aims at quantifying usability and security of gestures as well as understanding their relationship across multiple gestures. To study gesture-based authentication, we design an authentication method that combines Dynamic Time Warping (DTW) and Support Vector Machine (SVM), and conducted a user study with 42 participants over a period of 6 weeks. We objectively quantify the usability of a gesture by the number of corners and the frame length of all gesture samples, quantify the security using the equal error rate (EER), and the consistency by EER over a period of time. Meanwhile, we obtain subjective evaluation of usability and security by conducting a survey. By examining the responses, we found that the subjective evaluation confirms with the objective ones, and usability is in inverse relationship with security. We studied the consistency of gestures and found that most participants forgot gestures to some degree and reinforcing the memorization of gestures is necessary to improve the authentication performance. Finally, we performed a study with another 17 participants on shoulder surfing attacks, where attackers can observe the victims multiple times. The results show that shoulder surfing does not help to boost the attacks.

Xiaopeng Li, Xianshan Qu, Wenyuan Xu et al.

The contextual information (i.e., the time and location) in which a photo is taken can be easily tampered with or falsely claimed by forgers to achieve malicious purposes, e.g., creating fear among the general public. A rich body of work has focused on detecting photo tampering and manipulation by verifying the integrity of image content. Instead, we aim to detect photo misuse by verifying the capture time and location of photos. This paper is motivated by the law of nature that sun position varies with the time and location, which can be used to determine whether the claimed contextual information corresponds with the sun position that the image content actually indicates. Prior approaches to inferring sun position from images mainly rely on vanishing points associated with at least two shadows, while we propose novel algorithms which utilize only one shadow in the image to infer the sun position. Meanwhile, we compute the sun position by applying astronomical algorithms which take as input the claimed capture time and location. Only when the two estimated sun positions are consistent can the claimed contextual information be genuine. We have developed a prototype called IMAGEGUARD. The experimental results show that our method can successfully estimate sun position and detect the time-location inconsistency with high accuracy. By setting the thresholds to be 9.4 degrees and 5 degrees for the sun position distance and the altitude angle distance, respectively, our system can correctly identify 91.5% of falsified photos with fake contextual information.

Guoming Zhang, Chen Yan, Xiaoyu Ji et al.

Speech recognition (SR) systems such as Siri or Google Now have become an increasingly popular human-computer interaction method, and have turned various systems into voice controllable systems(VCS). Prior work on attacking VCS shows that the hidden voice commands that are incomprehensible to people can control the systems. Hidden voice commands, though hidden, are nonetheless audible. In this work, we design a completely inaudible attack, DolphinAttack, that modulates voice commands on ultrasonic carriers (e.g., f > 20 kHz) to achieve inaudibility. By leveraging the nonlinearity of the microphone circuits, the modulated low frequency audio commands can be successfully demodulated, recovered, and more importantly interpreted by the speech recognition systems. We validate DolphinAttack on popular speech recognition systems, including Siri, Google Now, Samsung S Voice, Huawei HiVoice, Cortana and Alexa. By injecting a sequence of inaudible voice commands, we show a few proof-of-concept attacks, which include activating Siri to initiate a FaceTime call on iPhone, activating Google Now to switch the phone to the airplane mode, and even manipulating the navigation system in an Audi automobile. We propose hardware and software defense solutions. We validate that it is feasible to detect DolphinAttack by classifying the audios using supported vector machine (SVM), and suggest to re-design voice controllable systems to be resilient to inaudible voice command attacks.