This paper examines how multi-order interactions among visual concepts impact the generalization performance of deepfake detectors. We categorize these interactions and apply Shapley values from game theory to fairly evaluate their contributions. We propose three hypotheses: 1. Excessive high-order interactions typically have a significant negative impact on the task during encoding by deep-fake detectors. 2. Deepfake detectors with strong generalization selectively encode ultra-high-order interactions that have relatively smaller negative impact. 3. Deepfake detectors typically mitigate negative impacts by reducing the intensity of ultra-high-order interactions. To verify these hypotheses, we designed metrics to evaluate the effects of low-, high-, and ultra-high-order interactions on detector performance. In experiments, the above hypotheses are verified among various models with different backbones. Based on our findings, we propose a universal method to enhance the generalization of the deepfake detector without retraining the model.

Deep Neural Networks (DNNs) are highly sus-ceptible to adversarial noise, which can lead to erroneous predictions. In high-stakes scenarios, such as autonomous driving and medical diagnosis, DNNs inaccuracies can be dire. To address this issue, Adversarial Training (AT) has been widely adopted as an effective defense method. However, our analysis reveals two critical flaws in the traditional AT approach that hinder its adversarial robustness: (1) focus only on a subset of robust features during the training process. This narrow focus limits the model's ability to learn and perceive a diverse range of features. (2) tend to overlook potential cues in non-robust features that could be beneficial for the model to make correct predictions. These cues, referred to as “positive activations” for simplicity, contain valuable information that can enhance the model's perception and understanding of the input data. In this way, we propose a novel and plug-and-play framework called Multi-branch Feature Fusion and Dynamic Reconstruction (MFFDR), which leverages multi-branch attention mechanisms to enhance the model's perception of robust features and enrich the diversity of learned features. Moreover, we employ a dynamic weighting strategy to reconstruct non-robust features in order to utilize the positive activations embedded within them. Extensive experiments demonstrate that our method significantly improves the model's adversarial robustness and outperforms previous state-of-the-art methods.

In recent years, with the rapid increase the popularity of cellular Internet of Things (IoT) devices and the sharp increase in the number of end users, ensuring the stability and reliability of IoT systems has become an important challenge. In this context, anomaly detection techniques provide solutions for IoT applications. This paper reviews anomaly detection research applied in the field of the IoT in recent years (mainly from 2018 to 2023) from a technical perspective. First, the causes and basic types of IoT anomalies are introduced to provide a better understanding of the importance of anomaly detection. Second, we focus on research progress in machine learning and edge computing, and propose a general workflow for anomaly detection in the IoT based on edge computing, which we call ECADW. Furthermore, the challenges of anomaly detection in the IoT are proposed, and future research per-spectives are prospected. It is hoped that this review can help researchers to better understand the research direction of this topic and choose interesting anomaly detection techniques.

Photovoltaic(PV) power generation forecasting is an important method to solve the inherent volatility and inter-mittency of solar energy. However, traditional research methods often assume that the data is complete or well pre-processed, which does not match the reality of data collection scenarios where data is missing. To this end, We propose M-VGTG, a novel framework that integrates Variational Mode Decompo-sition (VMD), Global-Local Temporal Attention Unit(GLTA-Unit), and Multi-Scale Temporal Graph Convolution to tackle the complex issue of missing values in PV power forecasting. First, to address the problem of modal aliasing, a missing data processing strategy based on VMD is designed. Secondly, a carefully designed GLTA-Unit provides a global-local attention mechanism to capture long-term dependencies and local fluctu-ations, improving the forecasting performance. To address the problem of different positive and negative correlation relation-ships between sequences at different time scales, a multi-scale method is adopted. The GLTA-Unit and Partial Convolution are introduced into the backbone model Temporal Convolutional Network (TCN) to handle temporal features and missing mode updates. Capturing spatial structure is also important. In order to adaptively handle missing modes, this paper uses Adaptive Graph Convolution Network (AGCN) to process spatial features and introduces the latest BiasedGCN module to handle the perception of missing values in the information propagation process. Experiments on real PV power generation datasets verify the effectiveness of the proposed method and demonstrate the potential of this architecture in addressing the unique challenges of PV power gencratlon forecasting.

In recent years, the use of Unmanned Aerial Vehicles (UAVs) equipped with Mobile Edge Computing (MEC) servers to provide computational resources to mobile devices(MDs) has emerged as a promising technology. This paper aims to investigate a UAV-assisted Mobile Edge Computing (MEC) system in dynamic scenarios with stochastic computing tasks. Our goal is to minimize the total energy consumption of MDs by optimizing user association, resource allocation, and UAV trajectory. Considering the nonconvexity of the problem and the coupling among variables, we propose a novel deep reinforcement learning algorithm called improved-DDPG. In this algorithm, we employ improved Prioritized Experience Replay (PER) to enhance the convergence of the training process, and we introduce the annealing concept to enhance the algorithm's exploration capability. Simulation results demonstrate that the improved-DDPG algorithm exhibits good convergence and stability. Compared to baseline approaches, the improved-DDPG algorithm effectively reduces the energy consumption of terminal devices.

Channel pruning has demonstrated its effectiveness in compressing ConvNets. In many related arts, the importance of an output feature map is only determined by its associated filter. However, these methods ignore a small part of weights in the next layer which disappears as the feature map is removed. They ignore the phenomenon of weight dependency. Besides, many pruning methods use only one criterion for evaluation and find a sweet spot of pruning structure and accuracy in a trial-and-error fashion, which can be time-consuming. In this paper, we proposed a channel pruning algorithm via multi-criteria based on weight dependency, CPMC, which can compress a pre-trained model directly. CPMC defines channel importance in three aspects, including its associated weight value, computational cost, and parameter quantity. According to the phenomenon of weight dependency, CPMC gets channel importance by assessing its associated filter and the corresponding partial weights in the next layer. Then CPMC uses global normalization to achieve cross-layer comparison. Finally, CPMC removes less important channels by global ranking. CPMC can compress various CNN models, including VGGNet, ResNet, and DenseNet on various image classification datasets. Extensive experiments have shown CPMC outperforms the others significantly.