Graph convolution networks (GCN) have been widely used in skeleton-based action recognition. We note that existing GCN-based approaches primarily rely on prescribed graphical structures (ie., a manually defined topology of skeleton joints), which limits their flexibility to capture complicated correlations between joints. To move beyond this limitation, we propose a new framework for skeleton-based action recognition, namely Dynamic Group Spatio-Temporal GCN (DG-STGCN). It consists of two modules, DG-GCN and DG-TCN, respectively, for spatial and temporal modeling. In particular, DG-GCN uses learned affinity matrices to capture dynamic graphical structures instead of relying on a prescribed one, while DG-TCN performs group-wise temporal convolutions with varying receptive fields and incorporates a dynamic joint-skeleton fusion module for adaptive multi-level temporal modeling. On a wide range of benchmarks, including NTURGB+D, Kinetics-Skeleton, BABEL, and Toyota SmartHome, DG-STGCN consistently outperforms state-of-the-art methods, often by a notable margin.
Deep learning models have achieved excellent recognition results on large-scale video benchmarks. However, they perform poorly when applied to videos with rare scenes or objects, primarily due to the bias of existing video datasets. We tackle this problem from two different angles: algorithm and dataset. From the perspective of algorithms, we propose Spatial-aware Multi-Aspect Debiasing (SMAD), which incorporates both explicit debiasing with multi-aspect adversarial training and implicit debiasing with the spatial actionness reweighting module, to learn a more generic representation invariant to non-action aspects. To neutralize the intrinsic dataset bias, we propose OmniDebias to leverage web data for joint training selectively, which can achieve higher performance with far fewer web data. To verify the effectiveness, we establish evaluation protocols and perform extensive experiments on both re-distributed splits of existing datasets and a new evaluation dataset focusing on the action with rare scenes. We also show that the debiased representation can generalize better when transferred to other datasets and tasks.
In unsupervised domain adaptation (UDA), directly adapting from the source to the target domain usually suffers significant discrepancies and leads to insufficient alignment. Thus, many UDA works attempt to vanish the domain gap gradually and softly via various intermediate spaces, dubbed domain bridging (DB). However, for dense prediction tasks such as domain adaptive semantic segmentation (DASS), existing solutions have mostly relied on rough style transfer and how to elegantly bridge domains is still under-explored. In this work, we resort to data mixing to establish a deliberated domain bridging (DDB) for DASS, through which the joint distributions of source and target domains are aligned and interacted with each in the intermediate space. At the heart of DDB lies a dual-path domain bridging step for generating two intermediate domains using the coarse-wise and the fine-wise data mixing techniques, alongside a cross-path knowledge distillation step for taking two complementary models trained on generated intermediate samples as 'teachers' to develop a superior 'student' in a multi-teacher distillation manner. These two optimization steps work in an alternating way and reinforce each other to give rise to DDB with strong adaptation power. Extensive experiments on adaptive segmentation tasks with different settings demonstrate that our DDB significantly outperforms state-of-the-art methods. Code is available at https://github.com/xiaoachen98/DDB.git.
Recent years have witnessed significant success in Self-Supervised Learning (SSL), which facilitates various downstream tasks. However, attackers may steal such SSL models and commercialize them for profit, making it crucial to protect their Intellectual Property (IP). Most existing IP protection solutions are designed for supervised learning models and cannot be used directly since they require that the models' downstream tasks and target labels be known and available during watermark embedding, which is not always possible in the domain of SSL. To address such a problem especially when downstream tasks are diverse and unknown during watermark embedding, we propose a novel black-box watermarking solution, named SSL-WM, for protecting the ownership of SSL models. SSL-WM maps watermarked inputs by the watermarked encoders into an invariant representation space, which causes any downstream classifiers to produce expected behavior, thus allowing the detection of embedded watermarks. We evaluate SSL-WM on numerous tasks, such as Computer Vision (CV) and Natural Language Processing (NLP), using different SSL models, including contrastive-based and generative-based. Experimental results demonstrate that SSL-WM can effectively verify the ownership of stolen SSL models in various downstream tasks. Furthermore, SSL-WM is robust against model fine-tuning and pruning attacks. Lastly, SSL-WM can also evade detection from evaluated watermark detection approaches, demonstrating its promising application in protecting the IP of SSL models.
In this study, we dive deep into the unique challenges in semi-supervised object detection~(SSOD). We observe that current detectors generally suffer from 3 inconsistency problems. 1) Assignment inconsistency, that the conventional assignment policy is sensitive to labeling noise. 2) Subtasks inconsistency, where the classification and regression predictions are misaligned at the same feature point. 3) Temporal inconsistency, that the pseudo bboxes vary dramatically at different training steps. These issues lead to inconsistent optimization objectives of the student network, thus deteriorating performance and slowing down the model convergence. We, therefore, propose a systematic solution, termed Consistent Teacher, to remedy the above-mentioned challenges. First, adaptive anchor assignment substitutes the static IoU-based strategy, which enables the student network to be resistant to noisy psudo bboxes; Then we calibrate the subtask predictions by designing a feature alignment module; Lastly, We adopt a Gaussian Mixture Model (GMM) to dynamically adjust the pseudo-boxes threshold. Consistent Teacher provides a new strong baseline on a large range of SSOD evaluations. It achieves 40.0 mAP with ResNet-50 backbone given only 10% of annotated MS-COCO data, which surpasses previous baselines using pseudo labels by around 4 mAP. When trained on fully annotated MS-COCO with additional unlabeled data, the performance further increases to 49.1 mAP. Our code will be open-sourced soon.
Deep neural network (DNN) based speech enhancement models have attracted extensive attention due to their promising performance. However, it is difficult to deploy a powerful DNN in real-time applications because of its high computational cost. Typical compression methods such as pruning and quantization do not make good use of the data characteristics. In this paper, we introduce the Skip-RNN strategy into speech enhancement models with parallel RNNs. The states of the RNNs update intermittently without interrupting the update of the output mask, which leads to significant reduction of computational load without evident audio artifacts. To better leverage the difference between the voice and the noise, we further regularize the skipping strategy with voice activity detection (VAD) guidance, saving more computational load. Experiments on a high-performance speech enhancement model, dual-path convolutional recurrent network (DPCRN), show the superiority of our strategy over strategies like network pruning or directly training a smaller model. We also validate the generalization of the proposed strategy on two other competitive speech enhancement models.
With the broad application of deep neural networks (DNNs), backdoor attacks have gradually attracted attention. Backdoor attacks are insidious, and poisoned models perform well on benign samples and are only triggered when given specific inputs, which cause the neural network to produce incorrect outputs. The state-of-the-art backdoor attack work is implemented by data poisoning, i.e., the attacker injects poisoned samples into the dataset, and the models trained with that dataset are infected with the backdoor. However, most of the triggers used in the current study are fixed patterns patched on a small fraction of an image and are often clearly mislabeled, which is easily detected by humans or defense methods such as Neural Cleanse and SentiNet. Also, it's difficult to be learned by DNNs without mislabeling, as they may ignore small patterns. In this paper, we propose a generalized backdoor attack method based on the frequency domain, which can implement backdoor implantation without mislabeling and accessing the training process. It is invisible to human beings and able to evade the commonly used defense methods. We evaluate our approach in the no-label and clean-label cases on three datasets (CIFAR-10, STL-10, and GTSRB) with two popular scenarios (self-supervised learning and supervised learning). The results show our approach can achieve a high attack success rate (above 90%) on all the tasks without significant performance degradation on main tasks. Also, we evaluate the bypass performance of our approach for different kinds of defenses, including the detection of training data (i.e., Activation Clustering), the preprocessing of inputs (i.e., Filtering), the detection of inputs (i.e., SentiNet), and the detection of models (i.e., Neural Cleanse). The experimental results demonstrate that our approach shows excellent robustness to such defenses.
The recently proposed semi-blind source separation (SBSS) method for nonlinear acoustic echo cancellation (NAEC) outperforms adaptive NAEC in attenuating the nonlinear acoustic echo. However, the multiplicative transfer function (MTF) approximation makes it unsuitable for real-time applications especially in highly reverberant environments, and the natural gradient makes it hard to balance well between fast convergence speed and stability. In this paper, we propose two more effective SBSS methods based on auxiliary-function-based independent vector analysis (AuxIVA) and independent low-rank matrix analysis (ILRMA). The convolutive transfer function (CTF) approximation is used instead of MTF so that a long impulse response can be modeled with a short latency. The optimization schemes used in AuxIVA and ILRMA are carefully regularized according to the constrained demixing matrix of NAEC. Experimental results validate significantly better echo cancellation performance of the proposed methods.
Vertical federated learning (VFL) is attracting much attention because it enables cross-silo data cooperation in a privacy-preserving manner. While most research works in VFL focus on linear and tree models, deep models (e.g., neural networks) are not well studied in VFL. In this paper, we focus on SplitNN, a well-known neural network framework in VFL, and identify a trade-off between data security and model performance in SplitNN. Briefly, SplitNN trains the model by exchanging gradients and transformed data. On the one hand, SplitNN suffers from the loss of model performance since multiply parties jointly train the model using transformed data instead of raw data, and a large amount of low-level feature information is discarded. On the other hand, a naive solution of increasing the model performance through aggregating at lower layers in SplitNN (i.e., the data is less transformed and more low-level feature is preserved) makes raw data vulnerable to inference attacks. To mitigate the above trade-off, we propose a new neural network protocol in VFL called Security Forward Aggregation (SFA). It changes the way of aggregating the transformed data and adopts removable masks to protect the raw data. Experiment results show that networks with SFA achieve both data security and high model performance.
The direct expansion of deep neural network (DNN) based wide-band speech enhancement (SE) to full-band processing faces the challenge of low frequency resolution in low frequency range, which would highly likely lead to deteriorated performance of the model. In this paper, we propose a learnable spectral compression mapping (SCM) to effectively compress the high frequency components so that they can be processed in a more efficient manner. By doing so, the model can pay more attention to low and middle frequency range, where most of the speech power is concentrated. Instead of suppressing noise in a single network structure, we first estimate a spectral magnitude mask, converting the speech to a high signal-to-ratio (SNR) state, and then utilize a subsequent model to further optimize the real and imaginary mask of the pre-enhanced signal. We conduct comprehensive experiments to validate the efficacy of the proposed method.