Binary code analysis is essential for software security across various instruction set architectures. Cross-architecture binary function similarity detection faces significant challenges due to substantial differences in instruction sets and architectural conventions. Existing approaches struggle to capture relationships between code abstraction levels, and lack comprehensive cross-architecture datasets for effective evaluation. Inspired by human cognitive processes of dynamically integrating multi-level information, we propose Binary Dynamic Layer Fusion (BDLF), a novel neural architecture that enhances cross-architecture similarity detection through adaptive layer-wise feature integration. BDLF leverages Qwen3's multilingual code understanding and introduces dynamic weight generation to optimally combine representations from all previous layers. We also construct Cross-Bin, a high quality cross-architecture binary function dataset. BDLF-Qwen3 employs two-stage training: partial fine-tuning with pairwise similarity learning followed by BDLF enhancement with InfoNCE contrastive learning. Experiments demonstrate BDLF-Qwen3 significantly outperforms state-of-the-art methods, achieving 36-65\% improvement in Recall@10 across diverse CPU architectures.

The automated generation of secure network measurement scripts is vital for reliable network operations, but traditional models often neglect security, leading to vulnerabilities like command injection and buffer overflows. We propose PTC (Prefix Tuning CodeT5), a novel method leveraging prefix tuning on the CodeT5 model to generate secure, functionally correct scripts without full retraining. PTC introduces a security-oriented loss function to enhance code correctness and security compliance, reducing vulnerabilities while meeting functional requirements. We evaluate PTC on several groups of targets, and results show that our solution outperforms baseline models, producing secure and functional network measurement scripts across various scenarios. This approach effectively bridges the gap between code generation and security, offering a practical solution for developing robust network tools for real-world applications.

Garment sewing patterns are fundamental design elements that bridge the gap between design concepts and practical manufacturing. The generative modeling of sewing patterns is crucial for creating diversified garments. However, existing approaches are limited either by reliance on a single input modality or by suboptimal generation efficiency. In this work, we present GarmentDiffusion, a new generative model capable of producing centimeter-precise, vectorized 3D sewing patterns from multimodal inputs (text, image, and incomplete sewing pattern). Our method efficiently encodes 3D sewing pattern parameters into compact edge token representations, achieving a sequence length that is 10 times shorter than that of the autoregressive SewingGPT in DressCode. By employing a diffusion transformer, we simultaneously denoise all edge tokens along the temporal axis, while maintaining a constant number of denoising steps regardless of dataset-specific edge and panel statistics. With all combination of designs of our model, the sewing pattern generation speed is accelerated by 100 times compared to SewingGPT. We achieve new state-of-the-art results on DressCodeData, as well as on the largest sewing pattern dataset, namely GarmentCodeData. The project website is available at https://shenfu-research.github.io/Garment-Diffusion.

Recovering call graphs of binary programs plays an instrumental role in facilitating inter-procedural analysis tasks and subsequent applications. A salient challenge inherent in this process is the identification of indirect call targets, i.e., indirect callees. Existing solutions all have high false positives and negatives, making call graphs inaccurate. In this paper, we introduce CALLEE, an approach combining transfer learning and contrastive learning. The key insight is that, deep neural networks (DNNs) can automatically identify patterns concerning indirect calls. Inspired by question-answering applications, we employ contrastive learning to answer the callsite-callee question. To overcome the data-intensive nature of DNNs, we use transfer learning to pre-train on easy-to-collect direct calls and then fine-tune with indirect calls. Upon evaluating CALLEE across various target sets, our findings underscored its efficacy, associating callsites with callees surpasses state-of-the-art solutions, achieving over seven to eight times higher performance in both MRR and Recall@5. Further, when implementing CALLEE within two distinct applications-binary code similarity detection and hybrid fuzzing-we observed a marked enhancement in their operational performance.

Assembly code search is vital for reducing the burden on reverse engineers, allowing them to quickly identify specific functions using natural language within vast binary programs.Despite its significance, this critical task is impeded by the complexities involved in building highquality datasets.This paper explores training a Large Language Model (LLM) to emulate a general compiler.By leveraging Ubuntu packages to compile a dataset of 20 billion tokens, we further continue pre-train CodeLlama as a Virtual Compiler (ViC), capable of compiling any source code of any language to assembly code.This approach allows for virtual compilation across a wide range of programming languages without the need for a real compiler, preserving semantic equivalency and expanding the possibilities for assembly code dataset construction.Furthermore, we use ViC to construct a sufficiently large dataset for assembly code search.Employing this extensive dataset, we achieve a substantial improvement in assembly code search performance, with our model surpassing the leading baseline by 26%.

Reinforcement learning (RL) has been successfully applied to tackle many complex decision-making tasks. However, a critical problem is the unsafe exploration when deploying classical model-free RL methods in safety-critical systems. In this paper, we focus on safe RL for autonomous racing, which is formulated as a multi-objective optimization problem (MOP). We propose a novel mechanism, named Safe Action Generation (SAG), to address the above challenge. The mechanism mainly consists of two modules: a risk prediction module and an action generation module. The risk prediction module monitors the state of the vehicle in real-time, and once the state is evaluated as risky, the action generation module is activated to prevent unsafe behavior by restricting the exploration of the RL agent. Extensive experiments in The Open Racing Car Simulator (TORCS) demonstrate that our approach can complete one lap at an average speed of 108 km/h, Compared to the baseline without the SAG mechanism, our approach reduces the instances of driving off the track by 97.7%. By using the proposed framework, the learned policy can improve vehicle safety and generalization performance.

Hypophonia is a common speech symptom related to Parkinson's disease, affecting human comprehension and effective communication. Unlike dysarthric speech, hypophonic speech is characterized by its low volume and breathy voice, which makes it challenging to be heard and understood by human and voice-controllable systems especially in noisy environments. Conventional speech enhancement techniques, primarily focusing on amplifying audio power or cancelling environmental noise, fall short in improving the intelligibility and perception for hypophonic speech. To enhance hypophonic speech, we present ClearAI, an innovative AI-powered technology to improve speech quality for individuals suffering from hypophonia. ClearAI first leverages voice conversion technology to create a parallel dataset composed of normal and corresponding hypophonic speech samples. Then, ClearAI incorporates a predictive model trained on augmented parallel data to estimate the optimal audio style from hypophonic speech to strengthen the audio intensity and enhance the speech patterns. Next, a speech restoration model is built on the generated parallel speech data to reconstruct clear speech from the style transferred speech. Our experimental results reveal that ClearAI leads to substantial improvements in audio intensity in both digital formats and over-the-air transmission. In addition, ClearAI successfully reduces the hypophonic speech recognition error rate by more than 30% in noisy environments. Our human test results also validate ClearAI enhanced speech has the best human perceptual quality compared with other baseline methods.

In this paper, we propose WavePurifier, an audio purification framework to defend against audio adversarial attacks. Audio adversarial attacks craft adversarial examples or perturbations to attack the automated speech recognition (ASR) models. Although existing defense mechanisms can detect such attacks and raise alarms, they fail to recover or maintain benign commands. Consequently, this leads to the denial of users' benign commands. Different than existing defenses, WavePurifier aims to purify adversarial examples, thereby rectifying the user's benign commands. We find that the forward diffusion process of the diffusion model effectively eliminates perturbations, whereas the reverse diffusion process restores benign speech. Based on this, we develop a hierarchical diffusion model to defend against audio adversarial examples. This model is capable of purifying different spectrogram bands to varying degrees. To validate the performance of WavePurifier, we purify the adversarial examples from 3 different adversarial attacks in 140 distinct settings. In total, we collect 78,864 diffused spectrograms and 21,000 purified audios. Then, we evaluate WavePurifier on 2 different ASR models, 4 commercial speech-to-text APIs, 2 real-world attack scenarios, and compare them against 7 existing defense approaches. Our result shows that WavePurifier is a universal framework, demonstrating adaptability across diverse attacks with the same hyperparameters. Notably, WavePurifier outperforms existing methods with the lowest character error rate (CER), word error rate (WER), and a high purification success rate against different attacks.

In this article, a reinforcement learning (RL)-based strategy for unmanned surface vehicle (USV) path following control is developed. The proposed method learns integrated guidance and heading control policy, which directly maps the USV's navigation states to motor control commands. By introducing a twin-critic design and an integral compensator to the conventional deep deterministic policy gradient (DDPG) algorithm, the tracking accuracy and robustness of the controller can be significantly improved. Moreover, a pretrained neural network-based USV model is built to help the learning algorithm efficiently deal with unknown nonlinear dynamics. The self-learning and path following capabilities of the proposed method were validated in both simulations and real sea experiments. The results show that our control policy can achieve better performance than a traditional cascade control policy and a DDPG-based control policy.

Recent advances in natural language processing and machine learning have led to the development of chatbot models, such as ChatGPT, that can engage in conversational dialogue with human users. However, understanding the ability of these models to generate toxic or harmful responses during a non-toxic multi-turn conversation remains an open research problem. Existing research focuses on single-turn sentence testing, while we find that 82% of the individual non-toxic sentences that elicit toxic behaviors in a conversation are considered safe by existing tools. In this paper, we design a new attack, ToxicChat, by fine-tuning a chatbot to engage in conversation with a target open-domain chatbot. The chatbot is fine-tuned with a collection of crafted conversation sequences. Particularly, each conversation begins with a sentence from a crafted prompt sentences dataset. Our extensive evaluation shows that open-domain chatbot models can be triggered to generate toxic responses in a multi-turn conversation. In the best scenario, ToxicChat achieves a 67% toxicity activation rate. The conversation sequences in the fine-tuning stage help trigger the toxicity in a conversation, which allows the attack to bypass two defense methods. Our findings suggest that further research is needed to address chatbot toxicity in a dynamic interactive environment. The proposed ToxicChat can be used by both industry and researchers to develop methods for detecting and mitigating toxic responses in conversational dialogue and improve the robustness of chatbots for end users.

In this paper, we propose PhantomSound, a query-efficient black-box attack toward voice assistants. Existing black-box adversarial attacks on voice assistants either apply substitution models or leverage the intermediate model output to estimate the gradients for crafting adversarial audio samples. However, these attack approaches require a significant amount of queries with a lengthy training stage. PhantomSound leverages the decision-based attack to produce effective adversarial audios, and reduces the number of queries by optimizing the gradient estimation. In the experiments, we perform our attack against 4 different speech-to-text APIs under 3 real-world scenarios to demonstrate the real-time attack impact. The results show that PhantomSound is practical and robust in attacking 5 popular commercial voice controllable devices over the air, and is able to bypass 3 liveness detection mechanisms with success rate. The benchmark result shows that PhantomSound can generate adversarial examples and launch the attack in a few minutes. We significantly enhance the query efficiency and reduce the cost of a successful untargeted and targeted adversarial attack by 93.1% and 65.5% compared with the state-of-the-art black-box attacks, using merely ∼ 300 queries (∼ 5 minutes) and ∼ 1,500 queries (∼ 25 minutes), respectively.

Deep learning based voice synthesis technology generates artificial human-like speeches, which has been used in deepfakes or identity theft attacks. Existing defense mechanisms inject subtle adversarial perturbations into the raw speech audios to mislead the voice synthesis models. However, optimizing the adversarial perturbation not only consumes substantial computation time, but it also requires the availability of entire speech. Therefore, they are not suitable for protecting live speech streams, such as voice messages or online meetings. In this paper, we propose VSMask, a real-time protection mechanism against voice synthesis attacks. Different from offline protection schemes, VSMask leverages a predictive neural network to forecast the most effective perturbation for the upcoming streaming speech. VSMask introduces a universal perturbation tailored for arbitrary speech input to shield a real-time speech in its entirety. To minimize the audio distortion within the protected speech, we implement a weight-based perturbation constraint to reduce the perceptibility of the added perturbation. We comprehensively evaluate VSMask protection performance under different scenarios. The experimental results indicate that VSMask can effectively defend against 3 popular voice synthesis models. None of the synthetic voice could deceive the speaker verification models or human ears with VSMask protection. In a physical world experiment, we demonstrate that VSMask successfully safeguards the real-time speech by injecting the perturbation over the air.

Existing approaches to constrained-input optimal control problems mainly focus on systems with input saturation, whereas other constraints, such as combined inequality constraints and state-dependent constraints, are seldom discussed. In this article, a reinforcement learning (RL)-based algorithm is developed for constrained-input optimal control of discrete-time (DT) systems. The deterministic policy gradient (DPG) is introduced to iteratively search the optimal solution to the Hamilton-Jacobi-Bellman (HJB) equation. To deal with input constraints, an action mapping (AM) mechanism is proposed. The objective of this mechanism is to transform the exploration space from the subspace generated by the given inequality constraints to the standard Cartesian product space, which can be searched effectively by existing algorithms. By using the proposed architecture, the learned policy can output control signals satisfying the given constraints, and the original reward function can be kept unchanged. In our study, the convergence analysis is given. It is shown that the iterative algorithm is convergent to the optimal solution of the HJB equation. In addition, the continuity of the iterative estimated Q -function is investigated. Two numerical examples are provided to demonstrate the effectiveness of our approach.

In this paper, we propose SpecPatch, a human-in-the loop adversarial audio attack on automated speech recognition (ASR) systems. Existing audio adversarial attacker assumes that the users cannot notice the adversarial audios, and hence allows the successful delivery of the crafted adversarial examples or perturbations. However, in a practical attack scenario, the users of intelligent voice-controlled systems (e.g., smartwatches, smart speakers, smartphones) have constant vigilance for suspicious voice, especially when they are delivering their voice commands. Once the user is alerted by a suspicious audio, they intend to correct the falsely-recognized commands by interrupting the adversarial audios and giving more powerful voice commands to overshadow the malicious voice. This makes the existing attacks ineffective in the typical scenario when the user's interaction and the delivery of adversarial audio coincide. To truly enable the imperceptible and robust adversarial attack and handle the possible arrival of user interruption, we design SpecPatch, a practical voice attack that uses a sub-second audio patch signal to deliver an attack command and utilize periodical noises to break down the communication between the user and ASR systems. We analyze the CTC (Connectionist Temporal Classification) loss forwarding and backwarding process and exploit the weakness of CTC to achieve our attack goal. Compared with the existing attacks, we extend the attack impact length (i.e., the length of attack target command) by 287%. Furthermore, we show that our attack achieves 100% success rate in both over-the-line and over-the-air scenarios amid user intervention.

Traffic measurement is key to many important network functions. Supporting real-time queries at the individual flow level over networkwide traffic represents a major challenge that has not been successfully addressed yet. This paper provides the first solutions in supporting real-time networkwide queries and allowing a local network function (for performance, security or management purpose) to make queries at any measurement point at any time on any flow’s networkwide statistics, while the packets of the flow may traverse different paths in the network, some of which may not come across the point where the query is made. Our trace-based experiments demonstrate that the proposed solutions significantly outperform the baseline solutions derived from the existing techniques.

Estimating the cardinality of a data stream is a fundamental problem underlying numerous applications such as traffic monitoring in a network or a datacenter, popularity tracking on social media, and cache optimization in proxy servers. Existing solutions suffer from high processing/query overhead or memory in-efficiency, which prevents them from operating online for data streams with very high arrival rates. This paper takes a new solution path different from the prior art and proposes a self-morphing bitmap, which combines operational simplicity with structural dynamics, allowing the bitmap to be morphed in a series of steps with an evolving sampling probability that automatically adapts to different stream sizes. We evaluate the self-morphing bitmap theoretically and experimentally. The results demonstrate that it significantly outperforms the prior art.

Smart home Internet of Things (IoT) has a vibrant market with a wide range of appliances and sensors, spanning across smart home, smart city, and smart factory. However, the security and privacy of these IoT systems have raised serious concerns. Currently, most IoT devices rely on electromagnetic wave-based radio frequency (RF) for communication. Yet, RF has several inherent limitations, such as shortage of spectrum, susceptible to interference, and vulnerable to eavesdropping or jamming attacks. This article presents URadio, a wideband ultrasonic communication system. By leveraging recent advances in reduced Graphene Oxide (rGO), we design a new type of electrostatic ultrasonic transducer, which can achieve more than$6\times $bandwidth than commercial ultrasonic transducers. With this new transducer, we design an OFDM communication system to maximize its data rate for smart home applications. We build a prototype of URadio on a wireless testbed and evaluate its performance in several real-world environments. Our experiments show that URadio can reach up to 360 kb/s data rate at a distance of 81 cm or 20 Kb/s data rate at a distance of 20 m, which supports a variety of smart home applications. We further showcase URadio’s resilience against eavesdropping and jamming attacks, as well as demonstrate its capability of securely localizing objects in an indoor environment.

Estimating the cardinality of a data stream is a fundamental problem underlying numerous applications such as traffic monitoring in a network or a datacenter and query optimization of Internet-scale P2P data networks. Existing solutions suffer from high processing/query overhead or memory in-efficiency, which prevents them from operating online for data streams with very high arrival rates. This paper takes a new solution path different from the prior art and proposes a self-morphing bitmap, which combines operational simplicity with structural dynamics, allowing the bitmap to be morphed in a series of steps with an evolving sampling probability that automatically adapts to different stream sizes. We further generalize the design of self-morphing bitmap. We evaluate the self-morphing bitmap theoretically and experimentally. The results demonstrate that it significantly outperforms the prior art.

Internet-connected devices have been surging rapidly during the past years. Many important applications based upon such devices have emerged, such as vehicle tracking systems. These applications often require real-time execution of a large number of computation tasks. Edge computing has shown great potential in processing frequent but less-demanding tasks. Additionally, cloud computing allows for great scalability when substantial computing resources are needed. Edge-cloud computing is a paradigm that combines edge computing and cloud computing. A key problem in edge-cloud computing is how to determine the execution location for each computation task. We propose a dynamic edge-twin computing model in the context of edge-cloud computing. It uses an evaluation mechanism to predict the completion times of the task at both an edge device and a cloud server. The completion time includes data transfer time and computing time and it is determined based on real-time information about the task and the computing environment. The task will be executed by the device with a shorter completion time. We have implemented a vehicle tracking system under the edge-twin model. The experimental results show that the edge-twin model outperforms edge-alone computing and cloud-alone computing.

Traffic measurement provides fundamental statistics for network management functions. To implement the measurement modules on the data plane for real-time query response, modern sketches are designed to work with limited on-die memory allocation from network processors and collect traffic statistics in epochs of a preset length. To handle real-time queries at arbitrary times over traffic in a preceding period${T}$(called${T}$-queries), the prior art sets the epoch length to${T \over n}$and keeps the measurement results in a window of$n - 1$past epochs to support approximate${T}$-queries. Such an approach however drastically increases the memory cost or decreases the accuracy in the query results if the memory allocation is fixed. In this paper, motivated by the concept of offloading in today's edge-cloud computing, we propose a collaborative edge-center traffic measurement model, where the traffic measurement modules at all network devices form the edge, which offloads the traffic measurement results to a measurement center possibly hosted in a datacenter. The center synthesizes the measurements from the past epochs and sends the aggregate results back to the measurement modules to support T-queries. We conduct experiments using real traffic traces to evaluate the performance of the proposed edge-center measurement model. The experimental results demonstrate that the proposed designs significantly outperform the prior art.

With the explosive increase in wireless devices, physical-layer signal analysis has become critically beneficial across distinctive domains including interference minimization in network planning, security and privacy (e.g., drone and spycam detection), and mobile health with remote sensing. While SDR is known to be highly effective in realizing such services, they are rarely deployed or used by the end-users due to the costly hardware ~1K USD (e.g., USRP). Low-cost SDRs (e.g., RTL-SDR) are available, but their bandwidth is limited to 2-3 MHz and operation range falls well below 2.4 GHz - the unlicensed band holding majority of the wireless devices. This paper presents SDR-Lite, the first zero-cost, software-only software defined radio (SDR) receiver that empowers commodity WiFi to retrieve the In-phase and Quadrature of an ambient signal. With the full compatibility to pervasively-deployed WiFi infrastructure (without any change to the hardware and firmware), SDR-Lite aims to spread the blessing of SDR receiver functionalities to billions of WiFi users and households to enhance our everyday lives. The key idea of SDR-Lite is to trick WiFi to begin packet reception (i.e., the decoding process) when the packet is absent, so that it accepts ambient signals in the air and outputs corresponding bits. The bits are then reconstructed to the original physical-layer waveform, on which diverse SDR applications are performed. Our comprehensive evaluation shows that the reconstructed signal closely reassembles the original ambient signal (>85% correlation). We extensively demonstrate SDR-Lite effectiveness across seven distinctive SDR receiver applications under three representative categories: (i) RF fingerprinting, (ii) spectrum monitoring, and (iii) (ZigBee) decoding. For instance, in security applications of drone and rogue WiFi AP detection, SDR-Lite achieves 99% and 97% accuracy, which is comparable to USRP.

In this paper, a deep reinforcement learning-based robust control strategy for quadrotor helicopters is proposed. The quadrotor is controlled by a learned neural network which directly maps the system states to control commands in an end-to-end style. The learning algorithm is developed based on the deterministic policy gradient algorithm. By introducing an integral compensator to the actor-critic structure, the tracking accuracy and robustness have been greatly enhanced. Moreover, a two-phase learning protocol which includes both offline and online learning phase is proposed for practical implementation. An offline policy is first learned based on a simplified quadrotor model. Then, the policy is online optimized in actual flight. The proposed approach is evaluated in the flight simulator. The results demonstrate that the offline learned policy is highly robust to model errors and external disturbances. It also shows that the online learning could significantly improve the control performance.

Middleboxes are widely used in modern net-works for a variety of network functions in cybersecurity, performance enhancement, and monitoring. Middlebox policy enforcement is however complex and tedious with unreliable manual re-configuration of legacy routers. The existing solution on automated policy enforcement relies on software-defined networking and does not apply to the traditional non-SDN net-works, which remain popular today in enterprise deployment and core networks. This paper proposes a new architecture based entirely on software-defined middleboxes (instead of using software-defined switches in the prior art) to enable dependable and automated policy enforcement in non-SDN networks whose routers forward packets based on traditional routing protocols that are not policy-sensitive. We present a hot-potato enforcement strategy, which is then enhanced with two optimizations for load-balanced policy enforcement. Further enhancements are made to relieve middlebox processing overhead and avoid packet fragmentation due to policy enforcement.

In this paper, we propose an end-to-end modular reinforcement learning architecture for a navigation task in complex dynamic environments with rapidly moving obstacles. In this architecture, the main task is divided into two subtasks: local obstacle avoidance and global navigation. For obstacle avoidance, we develop a two-stream Q-network, which processes spatial and temporal information separately and generates action values. The global navigation subtask is resolved by a conventional Q-network framework. An online learning network and an action scheduler are introduced to first combine two pretrained policies, and then continue exploring and optimizing until a stable policy is obtained. The two-stream Q-network obtains better performance than the conventional deep Q-learning approach in the obstacle avoidance subtask. Experiments on the main task demonstrate that the proposed architecture can efficiently avoid moving obstacles and complete the navigation task at a high success rate. The modular architecture enables parallel training and also demonstrates good generalization capability in different environments.