While text embeddings enable efficient semantic processing in LLMs, they remain vulnerable to inversion attacks that reconstruct sensitive original text. However, current defense methods typically treat text embeddings from the feature level independently, ignoring the exploitation of the mutual relation among the embedding construction pipeline. To address this limitation, we propose Eguard, a framework that effectively disrupts chains of relationships between the original semantic space and defended functional space. Our improvements manifest at two levels, i.e., the global-level and local-level mutual information. At the global level, we propose to minimize the statistical dependency between protected embeddings and their original inputs, effectively decoupling sensitive content from the semantic space accessible to adversaries. At the local level, we apply keyword-antonym contrastive learning to enforce semantic discriminability within the space of downstream utility. This synergy of global privacy control and local semantic alignment allows Eguard to achieve a superior privacy-utility trade-off than traditional defenses. Our approach significantly reduces privacy risks, protecting over 95 percent of tokens from inversion while maintaining high performance across downstream tasks consistent with original embeddings.

Abstract We present an efficient and scalable method for the inverse shape design problem of elastic objects, with broad applicability to diverse materials and interactive editing. The core idea is to decouple material nonlinearity from geometry optimization by introducing the Cauchy stress tensor as an auxiliary variable. We design a three‐stage scheme that iteratively optimizes the stress tensors and the rest shape, with each stage being well‐posed and efficiently‐solvable. To address the lack of a theoretical convergence guarantee arising from the decoupled energy formulation, we incorporate a relaxation method that ensures robust stability in practice. As a result, our method achieves a 3 × speedup over the state‐of‐the‐art asymptotic method [Jia21] on a model with 40k vertices and 112k elements (Fig. 2), and exhibits near‐linear scalability to large systems (Fig. 8). We demonstrate applications including rest shape design for various materials (ranging from standard models to complex spline‐based materials [XSZB15]), interactive material and force editing, and elastic object reconstruction from images.

The integration of microphones into a broad array of devices, from consumer electronics to industrial sensors, introduces vulnerabilities to out-of-band injection attacks, including ultrasound, laser, electromagnetic, and magnetic field attacks. These attacks enable adversaries to inject inaudible or imperceptible commands, compromising systems without direct physical access. This paper presents a robust, passive detection framework designed to address the full spectrum of out-of-band attacks on microphone-equipped devices. Unlike prior approaches, our system leverages advanced speech disentanglement to separate semantic and acoustic features from recorded audio, enabling a refined analysis of injection artifacts within each feature domain. By quantifying entropy-based chaos within the disen tangled representations, we detect subtle spectral and structural irregularities indicative of injected signals. The system further incorporates a preliminary stage to identify carrier traces where applicable, expediting detection in cases such as ultrasound and laser attacks. Extensive evaluations across various device types, including smartphones, tablets, and microphones, demonstrate the system's high accuracy and stability, achieving an AUC of 98% under diverse conditions and attack configurations.

Multimodal language models (LMs) have shown significant potential for applications across various domains but remain vulnerable to adversarial attacks. Current research in white-box or black-box settings generally struggles with unrealistic attack assumptions and limited efficacy of targeted attacks. This paper introduces CoGA, a novel gray-box collaborative adversarial attack method for multimodal LMs. Under our gray-box settings, attackers have access only to the victim model’s input encoders. With the guidance of different modalities, we perturb the embedding representations from encoders to disrupt the semantic alignment across modalities, ultimately causing inaccurate outputs on various downstream tasks. Specifically, we integrate text embeddings into the loss calculations of the image attack and utilize image embeddings to guide the ranking of vulnerable words and the selection of final samples. Extensive experiments demonstrate that our method achieves superior attack performance across diverse models and tasks, suggesting the shared vulnerability of multimodal LMs in confronting adversarial challenges. Our work provides new insights into the security of multimodal LMs, facilitating the deployment of more robust and secure models in practical applications.

Online voice communications are widely used nowadays. To protect speech from leakage, people tend to initiate the talk in sound-isolated environments. In this paper, we reveal a novel attack that recovers high-quality speech from outside soundproof zones. The rationale of the attack is to leverage sound-sensitive characteristics of piezoelectric materials, i.e., a piezo film that can change the phase of reflected mmWaves when placed in a sound field. If the attacker transmits mmWaves and analyzes reflected signals from the piezo film, the speech information can be compromised. More importantly, the piezo film is paper-like and works without a power supply. We propose a new speech recovery methodology to transform sound waves into wireless signals and build an end-to-end eavesdropping system working as a through-wall “microphone” to recover high-quality speech stealthily. To combat signal attenuation and improve speech quality, we develop a speech-enhancement scheme based on generative adversarial networks and propose to use multi-antenna information for intelligible speech reconstruction. We conduct extensive experiments to evaluate the system. The results indicate that the system achieves over 98% accuracy for digit recognition and works well over 5m away through the wall. We also test the system under complex scenarios and give countermeasures.

Device fingerprints can satisfy the high-security requirement of modern mobile applications (e.g., mobile payments) by guaranteeing the operation is performed on a trusted device. However, existing works on device fingerprints are weak to leakage, which leads to an irreversible failure of the device fingerprint authentication system after suffering from fingerprint theft attacks. The vulnerability drives us to propose a reconfigurable device fingerprint, i.e.,MotoPrint, that can recover the system after suffering from such attacks.MotoPrintstems from the motor vibration that can represent in both signals of the accelerometer and the gyroscope (i.e., they are homologous motion signals). Therefore, we designed a two-path feature extracting network and a sensor-independent training strategy to eliminate sensor noise that can decline authentication performance. In addition,MotoPrinthas a complete reconfiguration mechanism to cope with fingerprint leakage, which brings the damaged authentication system back to health. The evaluation of 80 stand-alone vibration motors and 20 in-built ones shows thatMotoPrintcan achieve high authentication accuracy of 98.5%. Meanwhile, we also demonstrate the reconfiguredMotoPrint, which can also effectively indicate the device's uniqueness with over 98% accuracy, is independent ofMotoPrints under other stimulating codes.

We propose MiNNIE, a simple yet comprehensive framework for real-time simulation of nonlinear near-incompressible elastics. To avoid the common volumetric locking issues at high Poisson's ratios of linear finite element methods (FEM), we build MiNNIE upon a mixed FEM framework and further incorporate a pressure stabilization term to ensure excellent convergence of multigrid solvers. Our pressure stabilization strategy injects bounded influence on nodal displacement which can be eliminated using a quasiNewton method. MiNNIE has a specially tailored GPU multigrid solver including a modified skinning-space interpolation scheme, a novel vertex Vanka smoother, and an efficient dense solver using Schur complement. MiNNIE supports various elastic material models and simulates them in real-time, supporting a full range of Poisson's ratios up to 0.5 while handling large deformations, element inversions, and self-collisions at the same time.

As automatic speech recognition evolves, deployment of the voice user interface (VUI) has boomingly expanded. Especially since the COVID-19 pandemic, the VUI has gained more attention in online communication owing to its non-contact property. However, the VUI struggles to be applied in public scenes due to the degradation of received audio signals caused by various ambient noises. In this article, we propose Wavoice , the first noise-resistant multi-modal speech recognition system that fuses two distinct voices sensing modalities (i.e., millimeter-wave signals and audio signals from a microphone) together. One key contribution is to model the inherent correlation between millimeter-wave and audio signals. Based on it, Wavoice facilitates the real-time noise-resistant voice activity detection and user targeting from multiple speakers. Additionally, we elaborate on two novel modules for multi-modal fusion embedded into the neural network, leading to accurate speech recognition. Extensive experiments prove the effectiveness of Wavoice under adverse conditions—that is, the character recognition error rate below 1% in a range of 7 m. In terms of robustness and accuracy, Wavoice considerably outperforms existing audio-only speech recognition methods with lower character error and word error rates.

An audio system containing loudspeakers and microphones is the fundamental hardware for voice-enabled devices, enabling voice interaction with mobile applications and smart homes. This paper presents MagBackdoor, the first magnetic field attack that injects malicious commands via a loudspeaker-based backdoor of the audio system, compromising the linked voice interaction system. MagBackdoor focuses on the magnetic threat on loudspeakers and manipulates their sound production stealthily. Consequently, the microphone will inevitably pick up malicious sound generated by the attacked speaker, due to the closely packed arrangement of internal audio systems. To prove the feasibility of MagBackdoor, we conduct comprehensive simulations and experiments. This study further models the mechanism by which an external magnetic field excites the sound production of loudspeakers, giving theoretical guidance to MagBackdoor. Aiming at stealthy magnetic attacks in real-world scenarios, we self-design a prototype that can emit magnetic fields modulated by voice commands. We implement MagBackdoor and evaluate it across a wide range of smart devices involving 16 smartphones, four laptops, two tablets, and three smart speakers, achieving an average 95% injection success rate with high-quality injected acoustic signals.

With the increasing deployment of voice-controlled devices in homes and enterprises, there is an urgent demand for voice identification to prevent unauthorized access to sensitive information and property loss. However, due to the broadcast nature of sound wave, a voice-only system is vulnerable to adverse conditions and malicious attacks. We observe that the cooperation of millimeter waves (mmWave) and voice signals can significantly improve the effectiveness and security of user identification. Based on the properties, we propose a multi-modal user identification system (named WavoID) by fusing the uniqueness of mmWave-sensed vocal vibration and mic-recorded voice of users. To estimate fine-grained waveforms, WavoID splits signals and adaptively combines useful decomposed signals according to correlative contents in both mmWave and voice. An elaborated anti-spoofing module in WavoID comprising biometric bimodal information defend against attacks. WavoID produces and fuses the response maps of mmWave and voice to improve the representation power of fused features, benefiting accurate identification, even facing adverse circumstances. We evaluate WavoID using commercial sensors on extensive experiments. WavoID has significant performance on user identification with over 98% accuracy on 100 user datasets.

More and more people turn to online voice communication with loudspeaker-equipped devices due to its convenience. To prevent speech leakage, soundproof rooms are often adopted. This paper presents mmPhone, a novel acoustic eavesdropping system that recovers loudspeaker speech protected by soundproof environments. The key idea is that properties of piezoelectric films in mmWave band can change with sound pressure due to the piezoelectric effect. If the property changes are acquired by an adversary (i.e., characterizing the piezoelectric effect with mmWaves), speech leakage can happen. More importantly, the piezoelectric film can work without a power supply. Base on this, we proposed a methodology using mmWaves to sense the film and decoding the speech from mmWaves, which turns the film into a passive "microphone". To recover intelligible speech, we further develop an enhancement scheme based on a denoising neural network, multi-channel augmentation, and speech synthesis, to compensate for the propagation and penetration loss of mmWaves. We perform extensive experiments to evaluate mmPhone and conduct digit recognition with over 93% accuracy. The results indicate mmPhone can recover high-quality and intelligible speech from a distance over 5m and is resilient to incident angles of sound waves (within 55 degrees) and different types of loudspeakers.

Earpiece mode of smartphones is often used for confidential communication. In this paper, we proposed a remote(>2m) and motion-resilient attack on smartphone earpiece. We developed an end-to-end eavesdropping system mmEve based on a commercial mmWave sensor to recover speech emitted from smartphone earpiece. The rationale of the attack is based on our observation that, soundwaves emitted from the smartphone's earpiece have a strong correlation with reflected mmWaves from the smartphone's rear. However, we find the recovered speech suffers from the sensor's self-noise and smartphone user's motion which limit attack distance to less than 2m, causing limited threats in real world. We modeled the motion interference under mmWave sensing and proposed a motion-resilient solution by optimizing the fitting function on I/Q plane. To achieve a practical attack with reasonable attack distance, we developed a GAN-based denoising scheme to eliminate the noise pattern of the sensor, which boosted the attack range to 6--8m. We evaluated mmEve with extensive experiments and find 23 different models of smartphones manufactured by Samsung, Huawei, etc. can be compromised by the proposed attack.

Meshes are an indispensable representation in many graphics applications because they provide conformal spatial discretizations. However, mesh-based operations are often slow due to unstructured memory access patterns. We propose MeshTaichi, a novel mesh compiler that provides an intuitive programming model for efficient mesh-based operations. Our programming model hides the complex indexing system from users and allows users to write mesh-based operations using reference-style neighborhood queries. Our compiler achieves its high performance by exploiting data locality. We partition input meshes and prepare the wanted relations by inspecting users' code during compile time. During run time, we further utilize on-chip memory (shared memory on GPU and L1 cache on CPU) to access the wanted attributes of mesh elements efficiently. Our compiler decouples low-level optimization options with computations, so that users can explore different localized data attributes and different memory orderings without changing their computation code. As a result, users can write concise code using our programming model to generate efficient mesh-based computations on both CPU and GPU backends. We test MeshTaichi on a variety of physically-based simulation and geometry processing applications with both triangle and tetrahedron meshes. MeshTaichi achieves a consistent speedup ranging from 1.4× to 6×, compared to state-of-the-art mesh data structures and compilers.

We propose a fast and robust solver to simulate continuum-based deformable models with constraints, in particular, rigid-body and joint constraints useful for soft articulated characters. Our method embeds the degrees of freedom of both articulated rigid bodies and deformable bodies in one unified constrained optimization problem, thus coupling the deformable and rigid bodies. Inspired by Projective Dynamics which is a fast numerical solver to simulate deformable objects, we also propose a novel local/global solver that takes full advantage of the pre-factorized system matrices to accelerate the solve of our constrained optimization problem. Therefore, our method can efficiently simulate character models, with rigid-body parts (bones) being correctly coupled with deformable parts (flesh). Our method is stable because backward Euler time integration is applied to both rigid and deformable degrees of freedom. Our unified optimization problem is rigorously derived from constrained Newtonian mechanics. When simulating only articulated rigid bodies as a special case, our method converges to the state-of-the-art rigid body simulators.

With the advance in automatic speech recognition, voice user interface has gained popularity recently. Since the COVID-19 pandemic, VUI is increasingly preferred in online communication due to its non-contact. Additionally, various ambient noise impedes the public applications of voice user interfaces due to the requirement of audio-only speech recognition methods for a high signal-to-noise ratio. In this paper, we present Wavoice, the first noise-resistant multi-modal speech recognition system that fuses two distinct voice sensing modalities, i.e., millimeter-wave (mmWave) signals and audio signals from a microphone, together. One key contribution is that we model the inherent correlation between mmWave and audio signals. Based on it, Wavoice facilitates the real-time noise-resistant voice activity detection and user targeting from multiple speakers. Furthermore, we elaborate on two novel modules into the neural attention mechanism for multi-modal signals fusion, and result in accurate speech recognition. Extensive experiments verify Wavoice's effectiveness under various conditions with the character recognition error rate below 1% in a range of 7 meters. Wavoice outperforms existing audio-only speech recognition methods with lower character error rate and word error rate. The evaluation in complex scenes validates the robustness of Wavoice.

We propose a new algorithm for updating a Cholesky factorization which speeds up Projective Dynamics simulations with topological changes. Our approach addresses an important limitation of the original Projective Dynamics, i.e., that topological changes such as cutting, fracturing, or tearing require full refactorization which compromises computation speed, especially in real-time applications. Our method progressively modifies the Cholesky factor of the system matrix in the global step instead of computing it from scratch. Only a small amount of overhead is added since most of the topological changes in typical simulations are continuous and gradual. Our method is based on the update and downdate routine in CHOLMOD, but unlike recent related work, supports dynamic sizes of the system matrix and the addition of new vertices. Our approach allows us to introduce clean cuts and perform interactive remeshing. Our experiments show that our method works particularly well in simulation scenarios involving cutting, tearing, and local remeshing operations.

We propose a simple yet efficient multigrid scheme to simulate high-resolution deformable objects in their full spaces at interactive frame rates. The point of departure of our method is the Galerkin projection which is simple to construct. However, a naïve Galerkin multigrid does not scale well for large and irregular grids because it trades-off matrix sparsity for smaller sized linear systems which eventually stops improving the performance. Given that observation, we design our special projection criterion which is based on skinning space coordinates with piecewise constant weights, to make our Galerkin multigrid method scale for high-resolution meshes without suffering from dense linear solves. The usage of skinning space coordinates enables us to reduce the resolution of grids more aggressively, and our piecewise constant weights further ensure us to always deal with reasonably-sparse linear solves. Our projection matrices also help us to manage multi-level linear systems efficiently. Therefore, our method can be applied to different optimization schemes such as Newton's method and Projective Dynamics, pushing the resolution of a real-time simulation to orders of magnitudes higher. Our final GPU implementation outperforms the other state-of-the-art GPU deformable body simulators, enabling us to simulate large deformable objects with hundred thousands of degrees of freedom in real-time.

We propose Average Vector Field (AVF) integration for simulation of deformable solids in physics-based animation. Our method achieves exact energy conservation for the St. Venant-Kirchhoff material without any correction steps or extra parameters. Exact energy conservation implies that our resulting animations 1) cannot explode and 2) do not suffer from numerical damping, which are two common problems with previous numerical integration techniques. Our method produces lively motion even with large time steps as typically used in physics-based animation. Our implicit update rules can be formulated as a minimization problem and solved in a similar way as optimization-based backward Euler, with only a mild computing overhead. Our approach also supports damping and collision response models, making it easy to deploy in practical computer animation pipelines.

We present a new time integration method featuring excellent stability and energy conservation properties, making it particularly suitable for real-time physics. The commonly used backward Euler method is stable but introduces artificial damping. Methods such as implicit midpoint do not suffer from artificial damping but are unstable in many common simulation scenarios. We propose an algorithm that blends between the implicit midpoint and forward/backward Euler integrators such that the resulting simulation is stable while introducing only minimal artificial damping. We achieve this by tracking the total energy of the simulated system, taking into account energy-changing events: damping and forcing. To facilitate real-time simulations, we propose a local/global solver, similar to Projective Dynamics, as an alternative to Newton’s method. Compared to the original Projective Dynamics, which is derived from backward Euler, our final method introduces much less numerical damping at the cost of minimal computing overhead. Stability guarantees of our method are derived from the stability of backward Euler, whose stability is a widely accepted empirical fact. However, to our knowledge, theoretical guarantees have so far only been proven for linear ODEs. We provide preliminary theoretical results proving the stability of backward Euler also for certain cases of nonlinear potential functions.

We propose a novel projection scheme that corrects energy fluctuations in simulations of deformable objects, thereby removing unwanted numerical dissipation and numerical "explosions". The key idea of our method is to first take a step using a conventional integrator, then project the result back to the constant energy-momentum manifold. We implement this strategy using fast projection , which only adds a small amount of overhead to existing physics-based solvers. We test our method with several implicit integration rules and demonstrate its benefits when used in conjunction with Position Based Dynamics and Projective Dynamics. When added to a dissipative integrator such as backward Euler, our method corrects the artificial damping and thus produces more vivid motion. Our projection scheme also effectively prevents instabilities that can arise due to approximate solves or large time steps. Our method is fast, stable, and easy to implement---traits that make it well-suited for real-time physics applications such as games or training simulators.

We present a new method for real-time physics-based simulation supporting many different types of hyperelastic materials. Previous methods such as Position-Based or Projective Dynamics are fast but support only a limited selection of materials; even classical materials such as the Neo-Hookean elasticity are not supported. Recently, Xu et al. [2015] introduced new “spline-based materials” that can be easily controlled by artists to achieve desired animation effects. Simulation of these types of materials currently relies on Newton’s method, which is slow, even with only one iteration per timestep. In this article, we show that Projective Dynamics can be interpreted as a quasi-Newton method. This insight enables very efficient simulation of a large class of hyperelastic materials, including the Neo-Hookean, spline-based materials, and others. The quasi-Newton interpretation also allows us to leverage ideas from numerical optimization. In particular, we show that our solver can be further accelerated using L-BFGS updates (Limited-memory Broyden-Fletcher-Goldfarb-Shanno algorithm). Our final method is typically more than 10 times faster than one iteration of Newton’s method without compromising quality. In fact, our result is often more accurate than the result obtained with one iteration of Newton’s method. Our method is also easier to implement, implying reduced software development costs.

Abstract We present a shape manipulation technique capable of producing deformations of 2D and 3D meshes, guaranteeing that no elements will be inverted. We achieve this by augmenting the quadratic ex‐rotated elastic energy with additional convex terms that penalize the presence of inverted elements. Using a schedule of increasing penalty coefficients, we efficiently and robustly converge to an inversion free state by solving a sequence of unconstrained convex minimization problems. This process can be interpreted as a special purpose Semi‐Definite Programming (SDP) solver. We demonstrate that our method outperforms solvers used in previous work, including commercial‐grade SDP software (MOSEK). As an additional benefit, our method also converges to the solution via a more intuitive path, which can be used for quick preview. We demonstrate the efficacy of our scheme in a number of 2D and 3D shapes undergoing moderate to drastic deformation.

We present a method to create personalized anatomical models ready for physics-based animation, using only a set of 3D surface scans. We start by building a template anatomical model of an average male which supports deformations due to both 1) subject-specific variations: shapes and sizes of bones, muscles, and adipose tissues and 2) skeletal poses. Next, we capture a set of 3D scans of an actor in various poses. Our key contribution is formulating and solving a large-scale optimization problem where we compute both subject-specific and pose-dependent parameters such that our resulting anatomical model explains the captured 3D scans as closely as possible. Compared to data-driven body modeling techniques that focus only on the surface, our approach has the advantage of creating physics-based models, which provide realistic 3D geometry of the bones and muscles, and naturally supports effects such as inertia, gravity, and collisions according to Newtonian dynamics.

We present a new method for implicit time integration of physical systems. Our approach builds a bridge between nodal Finite Element methods and Position Based Dynamics, leading to a simple, efficient, robust, yet accurate solver that supports many different types of constraints. We propose specially designed energy potentials that can be solved efficiently using an alternating optimization approach. Inspired by continuum mechanics, we derive a set of continuum-based potentials that can be efficiently incorporated within our solver. We demonstrate the generality and robustness of our approach in many different applications ranging from the simulation of solids, cloths, and shells, to example-based simulation. Comparisons to Newton-based and Position Based Dynamics solvers highlight the benefits of our formulation.

Characters with precise internal anatomy are important in film and visual effects, as well as in medical applications. We propose the first semi-automatic method for creating anatomical structures, such as bones, muscles, viscera and fat tissues. This is done by transferring a reference anatomical model from an input template to an arbitrary target character, only defined by its boundary representation (skin). The fat distribution of the target character needs to be specified. We can either infer this information from MRI data, or allow the users to express their creative intent through a new editing tool. The rest of our method runs automatically: it first transfers the bones to the target character, while maintaining their structure as much as possible. The bone layer, along with the target skin eroded using the fat thickness information, are then used to define a volume where we map the internal anatomy of the source model using harmonic (Laplacian) deformation. This way, we are able to quickly generate anatomical models for a large range of target characters, while maintaining anatomical constraints.

We describe a scheme for time integration of mass-spring systems that makes use of a solver based on block coordinate descent. This scheme provides a fast solution for classical linear (Hookean) springs. We express the widely used implicit Euler method as an energy minimization problem and introduce spring directions as auxiliary unknown variables. The system is globally linear in the node positions, and the non-linear terms involving the directions are strictly local. Because the global linear system does not depend on run-time state, the matrix can be pre-factored, allowing for very fast iterations. Our method converges to the same final result as would be obtained by solving the standard form of implicit Euler using Newton's method. Although the asymptotic convergence of Newton's method is faster than ours, the initial ratio of work to error reduction with our method is much faster than Newton's. For real-time visual applications, where speed and stability are more important than precision, we obtain visually acceptable results at a total cost per timestep that is only a fraction of that required for a single Newton iteration. When higher accuracy is required, our algorithm can be used to compute a good starting point for subsequent Newton's iteration.