计算机科学
卷积神经网络
Python(编程语言)
人工智能
规范化(社会学)
语音增强
实时计算
滤波器(信号处理)
深度学习
语音识别
移动设备
数据流挖掘
噪音(视频)
召回
目标检测
模态(人机交互)
音频信号
内存占用
降噪
变换编码
计算机视觉
定位关键字
帧(网络)
人工神经网络
音频信号处理
精确性和召回率
编码器
噪声测量
移动电话技术
作者
Jing Cao,Min Liu,Huan He,Xuefeng Xing
标识
DOI:10.1109/tce.2026.3677160
摘要
Nowadays, the rapid increase of deep fake technologies provides serious threats to the authenticity of real-time communications in consumer applications. Fake audio and video content can easily be injected into live calls, undermining user trust and compromising security. The research proposes a real-time communication security enhancement mechanism that detects and mitigates deep forgeries in audio-visual streams during video calls. Video and audio data are continuously captured and processed in small temporal chunks during live communication. A Wiener Filter (WF) is applied to the audio stream for noise reduction, while frame normalization is used to standardize the video input. To improve cross-modal analysis, a Modality Attention Fusion (MAF) strategy is used to dynamically weigh the importance of visual and auditory features. A novel 3D Convolutional Neural Network – Attention-Gated Recurrent Unit (3D-CNN-AGRU) model is employed to learn spatial-temporal and sequential features from the preprocessed data. The trained model is effortlessly embedded into the communication app, running silently in the background to monitor for potential deepfake intrusions. A 3D-CNN-AGRU demonstrated using python high accuracy (98.75%), recall (97.99%), precision (98.04%), specificity (97.20%), and F1-score (98.27%) when compared to existing methods. The suggested 3D-CNN-AGRU with MAF presents a practical way to improve real-time communication security.
科研通智能强力驱动
Strongly Powered by AbleSci AI