Investigating and Mitigating the Multimodal Hallucination Snowballing in Large Vision-Language Models

Though advanced in understanding visual information with human languages, Large Vision-Language Models (LVLMs) still suffer from multimodal hallucinations. A natural concern is that during multimodal interaction, the generated hallucinations could influence the LVLMs' subsequent generation. Thus, we raise a question: When presented with a query relevant to the previously generated hallucination, will LVLMs be misled and respond incorrectly, even though the ground visual information exists? To answer this, we propose a framework called MMHalSnowball to evaluate LVLMs' behaviors when encountering generated hallucinations, where LVLMs are required to answer specific visual questions within a curated hallucinatory conversation. Crucially, our experiment shows that the performance of open-source LVLMs drops by at least $31\%$, indicating that LVLMs are prone to accept the generated hallucinations and make false claims that they would not have supported without distractions. We term this phenomenon Multimodal Hallucination Snowballing. To mitigate this, we further propose a training-free method called Residual Visual Decoding, where we revise the output distribution of LVLMs with the one derived from the residual visual input, providing models with direct access to the visual information. Experiments show that our method can mitigate more than $24\%$ of the snowballed multimodal hallucination while maintaining capabilities.

Low-Complexity Beam Training for Multi-RIS-Assisted Multi-User Communications

In this paper, we investigate the beam training problem in the multi-user millimeter wave (mmWave) communication system, where multiple reconfigurable intelligent surfaces (RISs) are deployed to improve the coverage and the achievable rate. However, existing beam training techniques in mmWave systems suffer from the high complexity (i.e., exponential order) and low identification accuracy. To address these problems, we propose a novel hashing multi-arm beam (HMB) training scheme that reduces the training complexity to the logarithmic order with the high accuracy. Specifically, we first design a generation mechanism for HMB codebooks. Then, we propose a demultiplexing algorithm based on the soft decision to distinguish signals from different RIS reflective links. Finally, we utilize a multi-round voting mechanism to align the beams. Simulation results show that the proposed HMB training scheme enables simultaneous training for multiple RISs and multiple users, and reduces the beam training overhead to the logarithmic level. Moreover, it also shows that our proposed scheme can significantly improve the identification accuracy by at least 20% compared to existing beam training techniques.

Stochastic Geometry Analysis for Distributed RISs-Assisted mmWave Communications

Millimeter wave (mmWave) has attracted considerable attention due to its wide bandwidth and high frequency. However, it is highly susceptible to blockages, resulting in significant degradation of the coverage and the sum rate. A promising approach is deploying distributed reconfigurable intelligent surfaces (RISs), which can establish extra communication links. In this paper, we investigate the impact of distributed RISs on the coverage probability and the sum rate in mmWave wireless communication systems. Specifically, we first introduce the system model, which includes the blockage, the RIS and the user distribution models, leveraging the Poisson point process. Then, we define the association criterion and derive the conditional coverage probabilities for the two cases of direct association and reflective association through RISs. Finally, we combine the two cases using Campbell's theorem and the total probability theorem to obtain the closed-form expressions for the ergodic coverage probability and the sum rate. Simulation results validate the effectiveness of the proposed analytical approach, demonstrating that the deployment of distributed RISs significantly improves the ergodic coverage probability by 45.4% and the sum rate by over 1.5 times.

Hashing Beam Training for Near-Field Communications

In this paper, we investigate the millimeter-wave (mmWave) near-field beam training problem to find the correct beam direction. In order to address the high complexity and low identification accuracy of existing beam training techniques, we propose an efficient hashing multi-arm beam (HMB) training scheme for the near-field scenario. Specifically, we first design a set of sparse bases based on the polar domain sparsity of the near-field channel. Then, the random hash functions are chosen to construct the near-field multi-arm beam training codebook. Each multi-arm beam codeword is scanned in a time slot until all the predefined codewords are traversed. Finally, the soft decision and voting methods are applied to distinguish the signal from different base stations and obtain correctly aligned beams. Simulation results show that our proposed near-field HMB training method can reduce the beam training overhead to the logarithmic level, and achieve 96.4% identification accuracy of exhaustive beam training. Moreover, we also verify applicability under the far-field scenario.

Coverage and Rate Analysis for Distributed RISs-Assisted mmWave Communications

The millimeter wave (mmWave) has received considerable interest due to its expansive bandwidth and high frequency. However, a noteworthy challenge arises from its vulnerability to blockages, leading to reduced coverage and achievable rates. To address these limitations, a potential solution is to deploy distributed reconfigurable intelligent surfaces (RISs), which comprise many low-cost and passively reflected elements, and can facilitate the establishment of extra communication links. In this paper, we leverage stochastic geometry to investigate the ergodic coverage probability and the achievable rate in both distributed RISs-assisted single-cell and multi-cell mmWave wireless communication systems. Specifically, we first establish the system model considering the stochastically distributed blockages, RISs and users by the Poisson point process. Then we give the association criterion and derive the association probabilities, the distance distributions, and the conditional coverage probabilities for two cases of associations between base stations and users without or with RISs. Finally, we use Campbell's theorem and the total probability theorem to obtain the closed-form expressions of the ergodic coverage probability and the achievable rate. Simulation results verify the effectiveness of our analysis method, and demonstrate that by deploying distributed RISs, the ergodic coverage probability is significantly improved by approximately 50%, and the achievable rate is increased by more than 1.5 times.

Fast Beam Training for Near-Field Communication Systems

In millimeter-wave communications, large-scale antenna arrays are commonly employed to mitigate obstacle occlusion and path loss. However, these large-scale arrays generate pencil-shaped beams, which necessitate a higher number of training beams to cover the desired space. This results in the heavy beam training overhead. Furthermore, as the antenna aperture increases, users are more likely to be situated in the near-field region of the base station (BS) antenna array. This motivates our investigation into the beam training problem in the near-field region to achieve efficient beam alignment. To address the high complexity and low identification accuracy of existing beam training techniques, we propose an efficient hashing multi-arm beam (HMB) training scheme for the near-field scenario. Specifically, we first design a set of sparse bases based on the polar domain sparsity of the near-field channel and construct a near-field single-beam training codebook. Then, the hash functions are chosen to construct the near-field multi-arm beam training codebook. Each multi-arm beam training codeword is used in a time slot until the predefined codebook is traversed. Finally, the soft decision and voting methods are applied to distinguish the signal from different BS and obtain the correctly aligned beams. In addition, we provide the logically rigorous proof of computational complexity. Simulation results show that our proposed near-field HMB training method can achieve 96.4% identification accuracy of the exhaustive beam training method and greatly reduce the training overhead to the logarithmic level. Furthermore, we verify its applicability under the far-field scenario as well.

Eve Said Yes: AirBone Authentication for Head-Wearable Smart Voice Assistant

Recent advances in machine learning and natural language processing have fostered the enormous prosperity of smart voice assistants and their services, e.g., Alexa, Google Home, Siri, etc. However, voice spoofing attacks are deemed to be one of the major challenges of voice control security, and never stop evolving such as deep-learning-based voice conversion and speech synthesis techniques. To solve this problem outside the acoustic domain, we focus on head-wearable devices, such as earbuds and virtual reality (VR) headsets, which are feasible to continuously monitor the bone-conducted voice in the vibration domain. Specifically, we identify that air and bone conduction (AC/BC) from the same vocalization are coupled (or concurrent) and user-level unique, which makes them suitable behavior and biometric factors for multi-factor authentication (MFA). The legitimate user can defeat acoustic domain and even cross-domain spoofing samples with the proposed two-stage AirBone authentication. The first stage answers \textit{whether air and bone conduction utterances are time domain consistent (TC)} and the second stage runs \textit{bone conduction speaker recognition (BC-SR)}. The security level is hence increased for two reasons: (1) current acoustic attacks on smart voice assistants cannot affect bone conduction, which is in the vibration domain; (2) even for advanced cross-domain attacks, the unique bone conduction features can detect adversary's impersonation and machine-induced vibration. Finally, AirBone authentication has good usability (the same level as voice authentication) compared with traditional MFA and those specially designed to enhance smart voice security. Our experimental results show that the proposed AirBone authentication is usable and secure, and can be easily equipped by commercial off-the-shelf head wearables with good user experience.

Target-Grounded Graph-Aware Transformer for Aerial Vision-and-Dialog Navigation

This report details the methods of the winning entry of the AVDN Challenge in ICCV CLVL 2023. The competition addresses the Aerial Navigation from Dialog History (ANDH) task, which requires a drone agent to associate dialog history with aerial observations to reach the destination. For better cross-modal grounding abilities of the drone agent, we propose a Target-Grounded Graph-Aware Transformer (TG-GAT) framework. Concretely, TG-GAT first leverages a graph-aware transformer to capture spatiotemporal dependency, which benefits navigation state tracking and robust action planning. In addition,an auxiliary visual grounding task is devised to boost the agent's awareness of referred landmarks. Moreover, a hybrid augmentation strategy based on large language models is utilized to mitigate data scarcity limitations. Our TG-GAT framework won the AVDN Challenge, with 2.2% and 3.0% absolute improvements over the baseline on SPL and SR metrics, respectively. The code is available at https://github.com/yifeisu/TG-GAT.

Omnipotent Adversarial Training for Unknown Label-noisy and Imbalanced Datasets

Adversarial training is an important topic in robust deep learning, but the community lacks attention to its practical usage. In this paper, we aim to resolve a real-world application challenge, i.e., training a model on an imbalanced and noisy dataset to achieve high clean accuracy and robustness, with our proposed Omnipotent Adversarial Training (OAT). Our strategy consists of two innovative methodologies to address the label noise and data imbalance in the training set. We first introduce an oracle into the adversarial training process to help the model learn a correct data-label conditional distribution. This carefully-designed oracle can provide correct label annotations for adversarial training. We further propose logits adjustment adversarial training to overcome the data imbalance challenge, which can help the model learn a Bayes-optimal distribution. Our comprehensive evaluation results show that OAT outperforms other baselines by more than 20% clean accuracy improvement and 10% robust accuracy improvement under the complex combinations of data imbalance and label noise scenarios. The code can be found in https://github.com/GuanlinLee/OAT.

PND-Net: Physics based Non-local Dual-domain Network for Metal Artifact Reduction

Metal artifacts caused by the presence of metallic implants tremendously degrade the reconstructed computed tomography (CT) image quality, affecting clinical diagnosis or reducing the accuracy of organ delineation and dose calculation in radiotherapy. Recently, deep learning methods in sinogram and image domains have been rapidly applied on metal artifact reduction (MAR) task. The supervised dual-domain methods perform well on synthesized data, while unsupervised methods with unpaired data are more generalized on clinical data. However, most existing methods intend to restore the corrupted sinogram within metal trace, which essentially remove beam hardening artifacts but ignore other components of metal artifacts, such as scatter, non-linear partial volume effect and noise. In this paper, we mathematically derive a physical property of metal artifacts which is verified via Monte Carlo (MC) simulation and propose a novel physics based non-local dual-domain network (PND-Net) for MAR in CT imaging. Specifically, we design a novel non-local sinogram decomposition network (NSD-Net) to acquire the weighted artifact component, and an image restoration network (IR-Net) is proposed to reduce the residual and secondary artifacts in the image domain. To facilitate the generalization and robustness of our method on clinical CT images, we employ a trainable fusion network (F-Net) in the artifact synthesis path to achieve unpaired learning. Furthermore, we design an internal consistency loss to ensure the integrity of anatomical structures in the image domain, and introduce the linear interpolation sinogram as prior knowledge to guide sinogram decomposition. Extensive experiments on simulation and clinical data demonstrate that our method outperforms the state-of-the-art MAR methods.