Xun Chen

Object detection underpins safety-critical applications such as autonomous driving and intelligent surveillance, yet the inherent vulnerability of object detection models exposes them to the dual threat of adversarial and backdoor attacks. Existing defense methods are mostly passive, relying on prior knowledge of known attacks and thus failing to defend against unknown attacks. To address these problems, we introduce Dynamic Heterogeneous Redundancy (DHR) into object detection and build the Dynamic Heterogeneous Redundancy for robust Object Detection (DHR-OD) framework. DHR-OD defends against both threats through four modules, namely a heterogeneous executor pool, dynamic scheduling, multi-model arbitration, and closed-loop feedback, of which the multi-model arbitration module is the core. It fuses the box-level outputs of multiple heterogeneous detectors and identifies anomalies, turning the output differences among heterogeneous models into a defense. Experiments on COCO and PASCAL VOC show that DHR-OD retains 88.9% accuracy under PGD-20 attack and reduces the BadDet backdoor attack success rate from about 100% to about 35% without a dedicated detection module, making it a defense framework that covers both adversarial and backdoor threats while not relying on prior knowledge of specific attack forms.

Speech recognition technology is an important branch in the field of artificial intelligence, aiming to transform human speech into computer-readable text information. However, speech recognition technology still faces many challenges, such as noise interference, and accent and speech rate differences. An aim of this paper is to explore a deep learning-based speech recognition method to improve the accuracy and robustness of speech recognition. Firstly, this paper introduces the basic principles of speech recognition and existing mainstream technologies, and then focuses on the deep learning-based speech recognition method. Through comparative experiments, it is found that the self-attention mechanism performs best in speech recognition tasks. In order to further improve speech recognition performance, this paper proposes a deep learning model based on the self-attention mechanism with DCNN-GRU. The model realizes the dynamic attention to an input speech by introducing the self-attention mechanism in a neural network model instead of an RNN and with a deep convolutional neural network, which improves the robustness and recognition accuracy of this model. This experiment uses 170 h of Chinese dataset AISHELL-1. Compared with the deep convolutional neural network, the deep learning model based on the self-attention mechanism with DCNN-GRU accomplishes a reduction of at least 6% in CER. Compared with a bidirectional gated recurrent neural network, the deep learning model based on the self-attention mechanism with DCNN-GRU accomplishes a reduction of 0.7% in CER. And finally, this experiment is performed on a test set analyzed the influencing factors affecting the CER. The experimental results show that this model exhibits good performance in various noise environments and accent conditions.

Detecting small and densely packed objects in images remains a significant challenge in computer vision. Existing object detection methods often exhibit low accuracy and frequently miss detection when identifying dense small objects and require larger model parameters. This study introduces a novel detection framework designed to address these limitations by integrating advanced feature fusion and optimization techniques. Our approach focuses on enhancing both detection accuracy and parameter efficiency. The approach was evaluated on the open-source VisDrone2019 data set and compared with mainstream algorithms. Experimental results demonstrate a 70.2% reduction in network parameters and a 6.3% improvement in mAP@0.5 over the original YOLOv7 algorithm. These results demonstrate that the enhanced model surpasses existing algorithms in detecting small objects.

In the dynamic and uncertain stock market, precise forecasting and decision-making are crucial for profitability. Traditional deep neural networks (DNN) often struggle with capturing long-term dependencies and multi-scale features in complex financial time series data. To address these challenges, we introduce CMR-DQN, an innovative framework that integrates discrete wavelet transform (DWT) for multi-scale data analysis, temporal convolutional network (TCN) for extracting deep temporal features, and a GRU–LSTM–Attention mechanism to enhance the model’s focus and memory. Additionally, CMR-DQN employs the Rainbow DQN reinforcement learning strategy to learn optimal trading strategies in a simulated environment. CMR-DQN significantly improved the total return rate on six selected stocks, with increases ranging from 20.37% to 55.32%. It also demonstrated substantial improvements over the baseline model in terms of Sharpe ratio and maximum drawdown, indicating increased excess returns per unit of total risk and reduced investment risk. These results underscore the efficiency and effectiveness of CMR-DQN in handling multi-scale time series data and optimizing stock market decisions.

In this work, we consider a covert communication scenario, where a transmitter Alice communicates to a receiver Bob with the aid of a probabilistic and uninformed jammer against an adversary warden’s detection. The transmission status and power of the jammer are random and follow some priori probabilities. We first analyze the warden’s detection performance as a function of the jammer’s transmission probability, transmit power distribution, and Alice’s transmit power. We then maximize the covert throughput from Alice to Bob subject to a covertness constraint, by designing the covert communication strategies from three different perspectives: Alice’s perspective, the jammer’s perspective, and the global perspective. Our analysis reveals that the minimum jamming power should not always be zero in the probabilistic jamming strategy, which is different from that in the continuous jamming strategy presented in the literature. In addition, we prove that the minimum jamming power should be the same as Alice’s covert transmit power, depending on the covertness and average jamming power constraints. Furthermore, our results show that the probabilistic jamming can outperform the continuous jamming in terms of achieving a higher covert throughput under the same covertness and average jamming power constraints.

As excellent tools for aiding communication, an intelligent reflecting surface (IRS) and an unmanned aerial vehicle (UAV) can extend the coverage area, remove the blind area, and achieve a dramatic rate improvement. In this paper, we improve the secrecy rate (SR) performance of directional modulation (DM) networks using an IRS and UAV in combination. To fully explore the benefits of the IRS and UAV, two efficient methods are proposed to enhance the SR performance. The first approach computes the confidential message (CM) beamforming vector by maximizing the SR, and the signal-to-leakage-noise ratio (SLNR) method is used to optimize the IRS phase shift matrix (PSM), which is called Max-SR-SLNR. To reduce the computational complexity, the CM, artificial noise (AN) beamforming, and IRS phase shift design are independently designed in the following method. The CM beamforming vector is constructed based on the maximum ratio transmission (MRT) criteria along the channel from Alice-to-IRS, the AN beamforming vector is designed by null-space projection (NSP) on the remaining two channels, and the PSM of the IRS is directly given by the phase alignment (PA) method. This method is called the MRT-NSP-PA. The simulation results show that the SR performance of the Max-SR-SLNR method outperforms the MRT-NSP-PA method in the cases of small-scale and medium-scale IRSs, and the latter approaches the former in performance as the IRS tends to a larger scale.