计算机科学
人工智能
深度学习
预处理器
特征提取
计算机视觉
分类器(UML)
可扩展性
模式识别(心理学)
面子(社会学概念)
深层神经网络
面部识别系统
光学(聚焦)
卷积神经网络
护盾
人脸检测
稳健性(进化)
生物识别
棱锥(几何)
目标检测
仿真
图像扭曲
建筑
地标
特征(语言学)
人工神经网络
自编码
数字取证
作者
Aryan Upadhyay,Sujay Pandey,Utkarsh Kumar Pandey,Priyanshu Varshney,Shelly Gupta,Mukesh Kumar Tripathi
标识
DOI:10.1109/icsadl67539.2026.11452093
摘要
Recent developments in Deep Learning have made it frighteningly easy to create artificially realistic content, resulting in an increasingly depressing reality crisis on all digital media. These are systems that have been extracted from the lab and that have acquired social, political and other public uses in which they often are being misused, as in the case of fraud, disinformation and identity theft. The increasing risk of this threat further underscores the need for effective and automatic deepfake detection systems. To counter this threat, author herein present Deep Shield as a generic framework to recognise the AI-generated visual content. The system employs an MTCNN for precise face detection and alignment, and the classifier is trained to focus on facial regions of interest where artefacts are likely to occur. The extracted facial features are processed by a Meso-Net-based architecture designed to capture minor mesoscopic distortions and pixel-level warping artefacts arising from GAN-based synthesis. Unlike typical forensics techniques that get nullified under heavy compression, Deep Shield employs special preprocessing to preserve information at high frequencies. With training on a cross-dataset set based on Face Forensics++, the method shows strong generalisation to various resolutions and compression factors. Experimental results suggest that Deep Shield is capable of achieving better accuracy and efficiency than traditional classifiers, indicating the practicability and scalability of Deep Shield to real-time content verification and digital forensics.
科研通智能强力驱动
Strongly Powered by AbleSci AI