Fingerprints with similar morphological characteristics but from different individuals can lead to errors in individual identification, especially when dealing with large databases containing millions of fingerprints. To address this issue and enhance the accuracy of similar fingerprint identification, the use of the likelihood ratio (LR) model for quantitative evaluation of fingerprint evidence has emerged as an effective research method. In this study, the LR fingerprint evidence evaluation model was established by using mathematical statistical methods, such as parameter estimation and hypothesis testing. This involved various steps, including database construction, scoring, fitting, calculation, and visual evaluation. Under the same-source conditions, the optimal parameter methods selected by different number of minutiae are gamma and Weibull distribution, while normal, Weibull, and lognormal distributions were the fitting parameters selected for minutiae configurations. The fitting parameters selected by different number of minutiae under different-source conditions are lognormal distribution, and the parameter methods selected for different minutiae configurations include Weibull, gamma, and lognormal distributions. The results of the LR model showed increased accuracy as the number of minutiae increased, indicating strong discriminative and corrective power. However, the accuracy of the LR evaluation based on different configurations was comparatively lower. In addition, the LR models with different numbers of minutiae outperformed those with different minutiae configurations. Our study shows that the use of LR models based on parametric methods is favoured in reducing the risk of fingerprint evidence misidentification, improving the quantitative assessment methods of fingerprint evidence, and promoting fingerprint identification from experience to science.
目的探究智能生物痕迹荧光显现仪在现场接触类生物物证发现的应用效果.方法实验选取现场常见的接触类物证,经过智能生物痕迹荧光显现仪显现处理,利用激光照射,观察显现效果,最后进行DNA提取与检验.结果对于常见现场接触类物证,经智能生物痕迹荧光显现仪处理后能够有效显现出生物物证,荧光强度大,灵敏度较高,并不影响后续DNA检验.结论智能生物痕迹荧光显现仪对常见渗透性、非渗透性等多种客体上生物物证具有良好的显现效果,荧光显现后不影响后续DNA提取,对现场生物物证发现与提取具有较好的应用价值.
计算机视觉的快速发展推动了人脸伪造技术的显著进步,深度学习的快速发展也使得人脸伪造变得越来越简单且难以识别.深度伪造的泛滥给国家和社会造成了负面影响.针对现有的深度伪造检测方法存在的准确率低、鲁棒性差等问题,利用可学习的分片方法和通道注意力单元对视觉转换器进行改进,并利用改进后模型进行深度伪造检测.具体而言,将可变形分块模块引入模型的每个阶段(Stage),并将通道注意力和自注意力结合构成混合注意力单元.此方法在两个通用深度人脸提取数据集上达到了 0.981和0.989的曲线下面积(AUC),并取得了较好的鲁棒性和泛化性.
在当今数字时代下,上市公司财务造假具有专业性、隐蔽性、智能化特点,涉案数据体量大、数据类型多、数据关系复杂等是上市公司财务造假侦查面临的主要挑战.如何利用前沿人工智能(AI)技术对财务造假进行智能化数据挖掘和侦查是经济侦查亟待解决的问题.由于上市公司财务数据的不同特征字段之间存在潜在的依赖关系,为此提出了一种Transformer风格卷积神经网络模型(TSCNN)对上市公司的财务造假进行监测识别,该技术针对深度学习中流行的卷积神经网络普遍应用小卷积核,无法对全局特征依赖性进行建模等问题,利用Transform-er架构中自注意力机制捕获全局的特征依赖,同时嵌入的大卷积核卷积层编码特征权重,增强感知特征间依赖关系的能力.此外,TSCNN通过线性复杂度的Hadamard积简化自注意力机制的运算.研究表明,所提出的侦查技术在上市公司财务数据上的鉴别性能优于现有流行的ResNet和MLP等人工智能算法.该方法是人工智能技术在经济侦查应用中的探索,在创新升级传统经济侦查手段基础上进一步提升办案效率.
随着科学技术的迅速发展,基于深度学习生成的合成语音给语音认证系统和网络空间安全带来了新的挑战.针对现有检测模型准确率较低和语音特征挖掘不够充分的问题,提出了一种基于Involution算子和交叉注意力机制改进的合成语音检测方法.前端将语音数据提取线性频率倒谱系数(LFCC)特征和恒定Q变换(CQT)谱图特征,两个特征分别输入到后端的双分支网络中.后端网络使用ResNet18作为主干网络先进行浅层的特征学习,并将Involution算子嵌入主干网络,扩大特征图像学习区域,增强在空间范围内学习到的频谱图像特征信息.同时在训练分支之后引入cross-attention交叉注意力机制,使LFCC特征和CQT谱图特征构建交互的全局信息,强化模型对特征的深层挖掘.所提模型在ASVspoof2019 LA测试集上取得了 0.84％的等错误率和0.026的最小归一化串联检测代价函数的实验结果,展现了优于主流的检测模型.结果表明,改进的模型能够有效融合不同的频谱特征,提高模型的特征学习能力,从而强化模型的检测能力.
