Abstract:In recent years, extensive research has been conducted on the vulnerability of ASR systems, revealing that black-box adversarial example attacks pose significant threats to real-world ASR systems. However, most existing black-box attacks rely on queries to the target ASRs, which is impractical when queries are not permitted. In this paper, we propose ZQ-Attack, a transfer-based adversarial attack on ASR systems in the zero-query black-box setting. Through a comprehensive review and categorization of modern ASR technologies, we first meticulously select surrogate ASRs of diverse types to generate adversarial examples. Following this, ZQ-Attack initializes the adversarial perturbation with a scaled target command audio, rendering it relatively imperceptible while maintaining effectiveness. Subsequently, to achieve high transferability of adversarial perturbations, we propose a sequential ensemble optimization algorithm, which iteratively optimizes the adversarial perturbation on each surrogate model, leveraging collaborative information from other models. We conduct extensive experiments to evaluate ZQ-Attack. In the over-the-line setting, ZQ-Attack achieves a 100% success rate of attack (SRoA) with an average signal-to-noise ratio (SNR) of 21.91dB on 4 online speech recognition services, and attains an average SRoA of 100% and SNR of 19.67dB on 16 open-source ASRs. For commercial intelligent voice control devices, ZQ-Attack also achieves a 100% SRoA with an average SNR of 15.77dB in the over-the-air setting.
Abstract:Pedestrian trajectory prediction plays a pivotal role in ensuring the safety and efficiency of various applications, including autonomous vehicles and traffic management systems. This paper proposes a novel method for pedestrian trajectory prediction, called multi-stage goal-driven network (MGNet). Diverging from prior approaches relying on stepwise recursive prediction and the singular forecasting of a long-term goal, MGNet directs trajectory generation by forecasting intermediate stage goals, thereby reducing prediction errors. The network comprises three main components: a conditional variational autoencoder (CVAE), an attention module, and a multi-stage goal evaluator. Trajectories are encoded using conditional variational autoencoders to acquire knowledge about the approximate distribution of pedestrians' future trajectories, and combined with an attention mechanism to capture the temporal dependency between trajectory sequences. The pivotal module is the multi-stage goal evaluator, which utilizes the encoded feature vectors to predict intermediate goals, effectively minimizing cumulative errors in the recursive inference process. The effectiveness of MGNet is demonstrated through comprehensive experiments on the JAAD and PIE datasets. Comparative evaluations against state-of-the-art algorithms reveal significant performance improvements achieved by our proposed method.
Abstract:Foley audio, critical for enhancing the immersive experience in multimedia content, faces significant challenges in the AI-generated content (AIGC) landscape. Despite advancements in AIGC technologies for text and image generation, the foley audio dubbing remains rudimentary due to difficulties in cross-modal scene matching and content correlation. Current text-to-audio technology, which relies on detailed and acoustically relevant textual descriptions, falls short in practical video dubbing applications. Existing datasets like AudioSet, AudioCaps, Clotho, Sound-of-Story, and WavCaps do not fully meet the requirements for real-world foley audio dubbing task. To address this, we introduce the Multi-modal Image and Narrative Text Dubbing Dataset (MINT), designed to enhance mainstream dubbing tasks such as literary story audiobooks dubbing, image/silent video dubbing. Besides, to address the limitations of existing TTA technology in understanding and planning complex prompts, a Foley Audio Content Planning, Generation, and Alignment (CPGA) framework is proposed, which includes a content planning module leveraging large language models for complex multi-modal prompts comprehension. Additionally, the training process is optimized using Proximal Policy Optimization based reinforcement learning, significantly improving the alignment and auditory realism of generated foley audio. Experimental results demonstrate that our approach significantly advances the field of foley audio dubbing, providing robust solutions for the challenges of multi-modal dubbing. Even when utilizing the relatively lightweight GPT-2 model, our framework outperforms open-source multimodal large models such as LLaVA, DeepSeek-VL, and Moondream2. The dataset is available at https://github.com/borisfrb/MINT .
Abstract:Pre-training has emerged as a simple yet powerful methodology for representation learning across various domains. However, due to the expensive training cost and limited data, pre-training has not yet been extensively studied in correspondence pruning. To tackle these challenges, we propose a pre-training method to acquire a generic inliers-consistent representation by reconstructing masked correspondences, providing a strong initial representation for downstream tasks. Toward this objective, a modicum of true correspondences naturally serve as input, thus significantly reducing pre-training overhead. In practice, we introduce CorrMAE, an extension of the mask autoencoder framework tailored for the pre-training of correspondence pruning. CorrMAE involves two main phases, \ie correspondence learning and matching point reconstruction, guiding the reconstruction of masked correspondences through learning visible correspondence consistency. Herein, we employ a dual-branch structure with an ingenious positional encoding to reconstruct unordered and irregular correspondences. Also, a bi-level designed encoder is proposed for correspondence learning, which offers enhanced consistency learning capability and transferability. Extensive experiments have shown that the model pre-trained with our CorrMAE outperforms prior work on multiple challenging benchmarks. Meanwhile, our CorrMAE is primarily a task-driven pre-training method, and can achieve notable improvements for downstream tasks by pre-training on the targeted dataset. We hope this work can provide a starting point for correspondence pruning pre-training.
Abstract:Text-to-Audio (TTA) aims to generate audio that corresponds to the given text description, playing a crucial role in media production. The text descriptions in TTA datasets lack rich variations and diversity, resulting in a drop in TTA model performance when faced with complex text. To address this issue, we propose a method called Portable Plug-in Prompt Refiner, which utilizes rich knowledge about textual descriptions inherent in large language models to effectively enhance the robustness of TTA acoustic models without altering the acoustic training set. Furthermore, a Chain-of-Thought that mimics human verification is introduced to enhance the accuracy of audio descriptions, thereby improving the accuracy of generated content in practical applications. The experiments show that our method achieves a state-of-the-art Inception Score (IS) of 8.72, surpassing AudioGen, AudioLDM and Tango.
Abstract:Various threats posed by the progress in text-to-speech (TTS) have prompted the need to reliably trace synthesized speech. However, contemporary approaches to this task involve adding watermarks to the audio separately after generation, a process that hurts both speech quality and watermark imperceptibility. In addition, these approaches are limited in robustness and flexibility. To address these problems, we propose TraceableSpeech, a novel TTS model that directly generates watermarked speech, improving watermark imperceptibility and speech quality. Furthermore, We design the frame-wise imprinting and extraction of watermarks, achieving higher robustness against resplicing attacks and temporal flexibility in operation. Experimental results show that TraceableSpeech outperforms the strong baseline where VALL-E or HiFicodec individually uses WavMark in watermark imperceptibility, speech quality and resilience against resplicing attacks. It also can apply to speech of various durations.
Abstract:Graph neural networks are recognized for their strong performance across various applications, with the backpropagation algorithm playing a central role in the development of most GNN models. However, despite its effectiveness, BP has limitations that challenge its biological plausibility and affect the efficiency, scalability and parallelism of training neural networks for graph-based tasks. While several non-BP training algorithms, such as the direct feedback alignment, have been successfully applied to fully-connected and convolutional network components for handling Euclidean data, directly adapting these non-BP frameworks to manage non-Euclidean graph data in GNN models presents significant challenges. These challenges primarily arise from the violation of the i.i.d. assumption in graph data and the difficulty in accessing prediction errors for all samples (nodes) within the graph. To overcome these obstacles, in this paper we propose DFA-GNN, a novel forward learning framework tailored for GNNs with a case study of semi-supervised learning. The proposed method breaks the limitations of BP by using a dedicated forward training mechanism. Specifically, DFA-GNN extends the principles of DFA to adapt to graph data and unique architecture of GNNs, which incorporates the information of graph topology into the feedback links to accommodate the non-Euclidean characteristics of graph data. Additionally, for semi-supervised graph learning tasks, we developed a pseudo error generator that spreads residual errors from training data to create a pseudo error for each unlabeled node. These pseudo errors are then utilized to train GNNs using DFA. Extensive experiments on 10 public benchmarks reveal that our learning framework outperforms not only previous non-BP methods but also the standard BP methods, and it exhibits excellent robustness against various types of noise and attacks.
Abstract:Policy gradient methods have enabled deep reinforcement learning (RL) to approach challenging continuous control problems, even when the underlying systems involve highly nonlinear dynamics that generate complex non-smooth optimization landscapes. We develop a rigorous framework for understanding how policy gradient methods mollify non-smooth optimization landscapes to enable effective policy search, as well as the downside of it: while making the objective function smoother and easier to optimize, the stochastic objective deviates further from the original problem. We demonstrate the equivalence between policy gradient methods and solving backward heat equations. Following the ill-posedness of backward heat equations from PDE theory, we present a fundamental challenge to the use of policy gradient under stochasticity. Moreover, we make the connection between this limitation and the uncertainty principle in harmonic analysis to understand the effects of exploration with stochastic policies in RL. We also provide experimental results to illustrate both the positive and negative aspects of mollification effects in practice.
Abstract:Key-value~(KV) caching is an important technique to accelerate the inference of large language models~(LLMs), but incurs significant memory overhead. To compress the size of KV cache, existing methods often compromise precision or require extra data for calibration, limiting their practicality in LLM deployment. In this paper, we introduce \textbf{DecoQuant}, a novel data-free low-bit quantization technique based on tensor decomposition methods, to effectively compress KV cache. Our core idea is to adjust the outlier distribution of the original matrix by performing tensor decomposition, so that the quantization difficulties are migrated from the matrix to decomposed local tensors. Specially, we find that outliers mainly concentrate on small local tensors, while large tensors tend to have a narrower value range. Based on this finding, we propose to apply low-bit quantization to the large tensor, while maintaining high-precision representation for the small tensor. Furthermore, we utilize the proposed quantization method to compress the KV cache of LLMs to accelerate the inference and develop an efficient dequantization kernel tailored specifically for DecoQuant. Through extensive experiments, DecoQuant demonstrates remarkable efficiency gains, showcasing up to a $\sim$75\% reduction in memory footprint while maintaining comparable generation quality.
Abstract:The ambiguous appearance, tiny scale, and fine-grained classes of objects in remote sensing imagery inevitably lead to the noisy annotations in category labels of detection dataset. However, the effects and treatments of the label noises are underexplored in modern oriented remote sensing object detectors. To address this issue, we propose a robust oriented remote sensing object detection method through dynamic loss decay (DLD) mechanism, inspired by the two phase ``early-learning'' and ``memorization'' learning dynamics of deep neural networks on clean and noisy samples. To be specific, we first observe the end point of early learning phase termed as EL, after which the models begin to memorize the false labels that significantly degrade the detection accuracy. Secondly, under the guidance of the training indicator, the losses of each sample are ranked in descending order, and we adaptively decay the losses of the top K largest ones (bad samples) in the following epochs. Because these large losses are of high confidence to be calculated with wrong labels. Experimental results show that the method achieves excellent noise resistance performance tested on multiple public datasets such as HRSC2016 and DOTA-v1.0/v2.0 with synthetic category label noise. Our solution also has won the 2st place in the "fine-grained object detection based on sub-meter remote sensing imagery" track with noisy labels of 2023 National Big Data and Computing Intelligence Challenge.